1 | //===-- OpenMP.cpp -- Open MP directive lowering --------------------------===// |
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 | // Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/ |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "flang/Lower/OpenMP.h" |
14 | |
15 | #include "ClauseProcessor.h" |
16 | #include "Clauses.h" |
17 | #include "DataSharingProcessor.h" |
18 | #include "Decomposer.h" |
19 | #include "ReductionProcessor.h" |
20 | #include "Utils.h" |
21 | #include "flang/Common/idioms.h" |
22 | #include "flang/Lower/Bridge.h" |
23 | #include "flang/Lower/ConvertExpr.h" |
24 | #include "flang/Lower/ConvertVariable.h" |
25 | #include "flang/Lower/DirectivesCommon.h" |
26 | #include "flang/Lower/StatementContext.h" |
27 | #include "flang/Lower/SymbolMap.h" |
28 | #include "flang/Optimizer/Builder/BoxValue.h" |
29 | #include "flang/Optimizer/Builder/FIRBuilder.h" |
30 | #include "flang/Optimizer/Builder/Todo.h" |
31 | #include "flang/Optimizer/Dialect/FIRType.h" |
32 | #include "flang/Optimizer/HLFIR/HLFIROps.h" |
33 | #include "flang/Parser/characters.h" |
34 | #include "flang/Parser/parse-tree.h" |
35 | #include "flang/Semantics/openmp-directive-sets.h" |
36 | #include "flang/Semantics/tools.h" |
37 | #include "flang/Support/Flags.h" |
38 | #include "flang/Support/OpenMP-utils.h" |
39 | #include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" |
40 | #include "mlir/Dialect/OpenMP/OpenMPDialect.h" |
41 | #include "mlir/Transforms/RegionUtils.h" |
42 | #include "llvm/ADT/STLExtras.h" |
43 | #include "llvm/Frontend/OpenMP/OMPConstants.h" |
44 | #include "llvm/Support/CommandLine.h" |
45 | |
46 | using namespace Fortran::lower::omp; |
47 | using namespace Fortran::common::openmp; |
48 | |
49 | static llvm::cl::opt<bool> DumpAtomicAnalysis("fdebug-dump-atomic-analysis" ); |
50 | |
51 | //===----------------------------------------------------------------------===// |
52 | // Code generation helper functions |
53 | //===----------------------------------------------------------------------===// |
54 | |
55 | static void genOMPDispatch(lower::AbstractConverter &converter, |
56 | lower::SymMap &symTable, |
57 | semantics::SemanticsContext &semaCtx, |
58 | lower::pft::Evaluation &eval, mlir::Location loc, |
59 | const ConstructQueue &queue, |
60 | ConstructQueue::const_iterator item); |
61 | |
62 | static void processHostEvalClauses(lower::AbstractConverter &converter, |
63 | semantics::SemanticsContext &semaCtx, |
64 | lower::StatementContext &stmtCtx, |
65 | lower::pft::Evaluation &eval, |
66 | mlir::Location loc); |
67 | |
68 | namespace { |
69 | /// Structure holding information that is needed to pass host-evaluated |
70 | /// information to later lowering stages. |
71 | class HostEvalInfo { |
72 | public: |
73 | // Allow this function access to private members in order to initialize them. |
74 | friend void ::processHostEvalClauses(lower::AbstractConverter &, |
75 | semantics::SemanticsContext &, |
76 | lower::StatementContext &, |
77 | lower::pft::Evaluation &, |
78 | mlir::Location); |
79 | |
80 | /// Fill \c vars with values stored in \c ops. |
81 | /// |
82 | /// The order in which values are stored matches the one expected by \see |
83 | /// bindOperands(). |
84 | void collectValues(llvm::SmallVectorImpl<mlir::Value> &vars) const { |
85 | vars.append(ops.loopLowerBounds); |
86 | vars.append(ops.loopUpperBounds); |
87 | vars.append(ops.loopSteps); |
88 | |
89 | if (ops.numTeamsLower) |
90 | vars.push_back(ops.numTeamsLower); |
91 | |
92 | if (ops.numTeamsUpper) |
93 | vars.push_back(ops.numTeamsUpper); |
94 | |
95 | if (ops.numThreads) |
96 | vars.push_back(ops.numThreads); |
97 | |
98 | if (ops.threadLimit) |
99 | vars.push_back(ops.threadLimit); |
100 | } |
101 | |
102 | /// Update \c ops, replacing all values with the corresponding block argument |
103 | /// in \c args. |
104 | /// |
105 | /// The order in which values are stored in \c args is the same as the one |
106 | /// used by \see collectValues(). |
107 | void bindOperands(llvm::ArrayRef<mlir::BlockArgument> args) { |
108 | assert(args.size() == |
109 | ops.loopLowerBounds.size() + ops.loopUpperBounds.size() + |
110 | ops.loopSteps.size() + (ops.numTeamsLower ? 1 : 0) + |
111 | (ops.numTeamsUpper ? 1 : 0) + (ops.numThreads ? 1 : 0) + |
112 | (ops.threadLimit ? 1 : 0) && |
113 | "invalid block argument list" ); |
114 | int argIndex = 0; |
115 | for (size_t i = 0; i < ops.loopLowerBounds.size(); ++i) |
116 | ops.loopLowerBounds[i] = args[argIndex++]; |
117 | |
118 | for (size_t i = 0; i < ops.loopUpperBounds.size(); ++i) |
119 | ops.loopUpperBounds[i] = args[argIndex++]; |
120 | |
121 | for (size_t i = 0; i < ops.loopSteps.size(); ++i) |
122 | ops.loopSteps[i] = args[argIndex++]; |
123 | |
124 | if (ops.numTeamsLower) |
125 | ops.numTeamsLower = args[argIndex++]; |
126 | |
127 | if (ops.numTeamsUpper) |
128 | ops.numTeamsUpper = args[argIndex++]; |
129 | |
130 | if (ops.numThreads) |
131 | ops.numThreads = args[argIndex++]; |
132 | |
133 | if (ops.threadLimit) |
134 | ops.threadLimit = args[argIndex++]; |
135 | } |
136 | |
137 | /// Update \p clauseOps and \p ivOut with the corresponding host-evaluated |
138 | /// values and Fortran symbols, respectively, if they have already been |
139 | /// initialized but not yet applied. |
140 | /// |
141 | /// \returns whether an update was performed. If not, these clauses were not |
142 | /// evaluated in the host device. |
143 | bool apply(mlir::omp::LoopNestOperands &clauseOps, |
144 | llvm::SmallVectorImpl<const semantics::Symbol *> &ivOut) { |
145 | if (iv.empty() || loopNestApplied) { |
146 | loopNestApplied = true; |
147 | return false; |
148 | } |
149 | |
150 | loopNestApplied = true; |
151 | clauseOps.loopLowerBounds = ops.loopLowerBounds; |
152 | clauseOps.loopUpperBounds = ops.loopUpperBounds; |
153 | clauseOps.loopSteps = ops.loopSteps; |
154 | ivOut.append(RHS: iv); |
155 | return true; |
156 | } |
157 | |
158 | /// Update \p clauseOps with the corresponding host-evaluated values if they |
159 | /// have already been initialized but not yet applied. |
160 | /// |
161 | /// \returns whether an update was performed. If not, these clauses were not |
162 | /// evaluated in the host device. |
163 | bool apply(mlir::omp::ParallelOperands &clauseOps) { |
164 | if (!ops.numThreads || parallelApplied) { |
165 | parallelApplied = true; |
166 | return false; |
167 | } |
168 | |
169 | parallelApplied = true; |
170 | clauseOps.numThreads = ops.numThreads; |
171 | return true; |
172 | } |
173 | |
174 | /// Update \p clauseOps with the corresponding host-evaluated values if they |
175 | /// have already been initialized. |
176 | /// |
177 | /// \returns whether an update was performed. If not, these clauses were not |
178 | /// evaluated in the host device. |
179 | bool apply(mlir::omp::TeamsOperands &clauseOps) { |
180 | if (!ops.numTeamsLower && !ops.numTeamsUpper && !ops.threadLimit) |
181 | return false; |
182 | |
183 | clauseOps.numTeamsLower = ops.numTeamsLower; |
184 | clauseOps.numTeamsUpper = ops.numTeamsUpper; |
185 | clauseOps.threadLimit = ops.threadLimit; |
186 | return true; |
187 | } |
188 | |
189 | private: |
190 | mlir::omp::HostEvaluatedOperands ops; |
191 | llvm::SmallVector<const semantics::Symbol *> iv; |
192 | bool loopNestApplied = false, parallelApplied = false; |
193 | }; |
194 | } // namespace |
195 | |
196 | /// Stack of \see HostEvalInfo to represent the current nest of \c omp.target |
197 | /// operations being created. |
198 | /// |
199 | /// The current implementation prevents nested 'target' regions from breaking |
200 | /// the handling of the outer region by keeping a stack of information |
201 | /// structures, but it will probably still require some further work to support |
202 | /// reverse offloading. |
203 | static llvm::SmallVector<HostEvalInfo, 0> hostEvalInfo; |
204 | |
205 | /// Bind symbols to their corresponding entry block arguments. |
206 | /// |
207 | /// The binding will be performed inside of the current block, which does not |
208 | /// necessarily have to be part of the operation for which the binding is done. |
209 | /// However, block arguments must be accessible. This enables controlling the |
210 | /// insertion point of any new MLIR operations related to the binding of |
211 | /// arguments of a loop wrapper operation. |
212 | /// |
213 | /// \param [in] converter - PFT to MLIR conversion interface. |
214 | /// \param [in] op - owner operation of the block arguments to bind. |
215 | /// \param [in] args - entry block arguments information for the given |
216 | /// operation. |
217 | static void bindEntryBlockArgs(lower::AbstractConverter &converter, |
218 | mlir::omp::BlockArgOpenMPOpInterface op, |
219 | const EntryBlockArgs &args) { |
220 | assert(op != nullptr && "invalid block argument-defining operation" ); |
221 | assert(args.isValid() && "invalid args" ); |
222 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
223 | |
224 | auto bindSingleMapLike = [&converter](const semantics::Symbol &sym, |
225 | const mlir::BlockArgument &arg) { |
226 | fir::ExtendedValue extVal = converter.getSymbolExtendedValue(sym); |
227 | auto refType = mlir::dyn_cast<fir::ReferenceType>(arg.getType()); |
228 | if (refType && fir::isa_builtin_cptr_type(refType.getElementType())) { |
229 | converter.bindSymbol(sym, arg); |
230 | } else { |
231 | extVal.match( |
232 | [&](const fir::BoxValue &v) { |
233 | converter.bindSymbol(sym, fir::BoxValue(arg, v.getLBounds(), |
234 | v.getExplicitParameters(), |
235 | v.getExplicitExtents())); |
236 | }, |
237 | [&](const fir::MutableBoxValue &v) { |
238 | converter.bindSymbol( |
239 | sym, fir::MutableBoxValue(arg, v.getLBounds(), |
240 | v.getMutableProperties())); |
241 | }, |
242 | [&](const fir::ArrayBoxValue &v) { |
243 | converter.bindSymbol(sym, fir::ArrayBoxValue(arg, v.getExtents(), |
244 | v.getLBounds(), |
245 | v.getSourceBox())); |
246 | }, |
247 | [&](const fir::CharArrayBoxValue &v) { |
248 | converter.bindSymbol(sym, fir::CharArrayBoxValue(arg, v.getLen(), |
249 | v.getExtents(), |
250 | v.getLBounds())); |
251 | }, |
252 | [&](const fir::CharBoxValue &v) { |
253 | converter.bindSymbol(sym, fir::CharBoxValue(arg, v.getLen())); |
254 | }, |
255 | [&](const fir::UnboxedValue &v) { converter.bindSymbol(sym, arg); }, |
256 | [&](const auto &) { |
257 | TODO(converter.getCurrentLocation(), |
258 | "target map clause operand unsupported type" ); |
259 | }); |
260 | } |
261 | }; |
262 | |
263 | auto bindMapLike = |
264 | [&bindSingleMapLike](llvm::ArrayRef<const semantics::Symbol *> syms, |
265 | llvm::ArrayRef<mlir::BlockArgument> args) { |
266 | // Structure component symbols don't have bindings, and can only be |
267 | // explicitly mapped individually. If a member is captured implicitly |
268 | // we map the entirety of the derived type when we find its symbol. |
269 | llvm::SmallVector<const semantics::Symbol *> processedSyms; |
270 | llvm::copy_if(Range&: syms, Out: std::back_inserter(x&: processedSyms), |
271 | P: [](auto *sym) { return !sym->owner().IsDerivedType(); }); |
272 | |
273 | for (auto [sym, arg] : llvm::zip_equal(processedSyms, args)) |
274 | bindSingleMapLike(*sym, arg); |
275 | }; |
276 | |
277 | auto bindPrivateLike = [&converter, &firOpBuilder]( |
278 | llvm::ArrayRef<const semantics::Symbol *> syms, |
279 | llvm::ArrayRef<mlir::Value> vars, |
280 | llvm::ArrayRef<mlir::BlockArgument> args) { |
281 | llvm::SmallVector<const semantics::Symbol *> processedSyms; |
282 | for (auto *sym : syms) { |
283 | if (const auto *commonDet = |
284 | sym->detailsIf<semantics::CommonBlockDetails>()) { |
285 | llvm::transform(commonDet->objects(), std::back_inserter(x&: processedSyms), |
286 | [&](const auto &mem) { return &*mem; }); |
287 | } else { |
288 | processedSyms.push_back(Elt: sym); |
289 | } |
290 | } |
291 | |
292 | for (auto [sym, var, arg] : llvm::zip_equal(processedSyms, vars, args)) |
293 | converter.bindSymbol( |
294 | *sym, |
295 | hlfir::translateToExtendedValue( |
296 | var.getLoc(), firOpBuilder, hlfir::Entity{arg}, |
297 | /*contiguousHint=*/ |
298 | evaluate::IsSimplyContiguous(*sym, converter.getFoldingContext())) |
299 | .first); |
300 | }; |
301 | |
302 | // Process in clause name alphabetical order to match block arguments order. |
303 | // Do not bind host_eval variables because they cannot be used inside of the |
304 | // corresponding region, except for very specific cases handled separately. |
305 | bindMapLike(args.hasDeviceAddr.syms, op.getHasDeviceAddrBlockArgs()); |
306 | bindPrivateLike(args.inReduction.syms, args.inReduction.vars, |
307 | op.getInReductionBlockArgs()); |
308 | bindMapLike(args.map.syms, op.getMapBlockArgs()); |
309 | bindPrivateLike(args.priv.syms, args.priv.vars, op.getPrivateBlockArgs()); |
310 | bindPrivateLike(args.reduction.syms, args.reduction.vars, |
311 | op.getReductionBlockArgs()); |
312 | bindPrivateLike(args.taskReduction.syms, args.taskReduction.vars, |
313 | op.getTaskReductionBlockArgs()); |
314 | bindMapLike(args.useDeviceAddr.syms, op.getUseDeviceAddrBlockArgs()); |
315 | bindMapLike(args.useDevicePtr.syms, op.getUseDevicePtrBlockArgs()); |
316 | } |
317 | |
318 | /// Get the list of base values that the specified map-like variables point to. |
319 | /// |
320 | /// This function must be kept in sync with changes to the `createMapInfoOp` |
321 | /// utility function, since it must take into account the potential introduction |
322 | /// of levels of indirection (i.e. intermediate ops). |
323 | /// |
324 | /// \param [in] vars - list of values passed to map-like clauses, returned |
325 | /// by an `omp.map.info` operation. |
326 | /// \param [out] baseOps - populated with the `var_ptr` values of the |
327 | /// corresponding defining operations. |
328 | static void |
329 | (llvm::ArrayRef<mlir::Value> vars, |
330 | llvm::SmallVectorImpl<mlir::Value> &baseOps) { |
331 | llvm::transform(vars, std::back_inserter(baseOps), [](mlir::Value map) { |
332 | auto mapInfo = map.getDefiningOp<mlir::omp::MapInfoOp>(); |
333 | assert(mapInfo && "expected all map vars to be defined by omp.map.info" ); |
334 | |
335 | mlir::Value varPtr = mapInfo.getVarPtr(); |
336 | if (auto boxAddr = varPtr.getDefiningOp<fir::BoxAddrOp>()) |
337 | return boxAddr.getVal(); |
338 | |
339 | return varPtr; |
340 | }); |
341 | } |
342 | |
343 | /// Get the directive enumeration value corresponding to the given OpenMP |
344 | /// construct PFT node. |
345 | llvm::omp::Directive |
346 | (const parser::OpenMPConstruct &ompConstruct) { |
347 | return common::visit( |
348 | common::visitors{ |
349 | [](const parser::OpenMPAllocatorsConstruct &c) { |
350 | return llvm::omp::OMPD_allocators; |
351 | }, |
352 | [](const parser::OpenMPAssumeConstruct &c) { |
353 | return llvm::omp::OMPD_assume; |
354 | }, |
355 | [](const parser::OpenMPAtomicConstruct &c) { |
356 | return llvm::omp::OMPD_atomic; |
357 | }, |
358 | [](const parser::OpenMPBlockConstruct &c) { |
359 | return std::get<parser::OmpBlockDirective>( |
360 | std::get<parser::OmpBeginBlockDirective>(c.t).t) |
361 | .v; |
362 | }, |
363 | [](const parser::OpenMPCriticalConstruct &c) { |
364 | return llvm::omp::OMPD_critical; |
365 | }, |
366 | [](const parser::OpenMPDeclarativeAllocate &c) { |
367 | return llvm::omp::OMPD_allocate; |
368 | }, |
369 | [](const parser::OpenMPDispatchConstruct &c) { |
370 | return llvm::omp::OMPD_dispatch; |
371 | }, |
372 | [](const parser::OpenMPExecutableAllocate &c) { |
373 | return llvm::omp::OMPD_allocate; |
374 | }, |
375 | [](const parser::OpenMPLoopConstruct &c) { |
376 | return std::get<parser::OmpLoopDirective>( |
377 | std::get<parser::OmpBeginLoopDirective>(c.t).t) |
378 | .v; |
379 | }, |
380 | [](const parser::OpenMPSectionConstruct &c) { |
381 | return llvm::omp::OMPD_section; |
382 | }, |
383 | [](const parser::OpenMPSectionsConstruct &c) { |
384 | return std::get<parser::OmpSectionsDirective>( |
385 | std::get<parser::OmpBeginSectionsDirective>(c.t).t) |
386 | .v; |
387 | }, |
388 | [](const parser::OpenMPStandaloneConstruct &c) { |
389 | return common::visit( |
390 | common::visitors{ |
391 | [](const parser::OpenMPSimpleStandaloneConstruct &c) { |
392 | return c.v.DirId(); |
393 | }, |
394 | [](const parser::OpenMPFlushConstruct &c) { |
395 | return llvm::omp::OMPD_flush; |
396 | }, |
397 | [](const parser::OpenMPCancelConstruct &c) { |
398 | return llvm::omp::OMPD_cancel; |
399 | }, |
400 | [](const parser::OpenMPCancellationPointConstruct &c) { |
401 | return llvm::omp::OMPD_cancellation_point; |
402 | }, |
403 | [](const parser::OmpMetadirectiveDirective &c) { |
404 | return llvm::omp::OMPD_metadirective; |
405 | }, |
406 | [](const parser::OpenMPDepobjConstruct &c) { |
407 | return llvm::omp::OMPD_depobj; |
408 | }, |
409 | [](const parser::OpenMPInteropConstruct &c) { |
410 | return llvm::omp::OMPD_interop; |
411 | }}, |
412 | c.u); |
413 | }, |
414 | [](const parser::OpenMPUtilityConstruct &c) { |
415 | return common::visit( |
416 | common::visitors{[](const parser::OmpErrorDirective &c) { |
417 | return llvm::omp::OMPD_error; |
418 | }, |
419 | [](const parser::OmpNothingDirective &c) { |
420 | return llvm::omp::OMPD_nothing; |
421 | }}, |
422 | c.u); |
423 | }}, |
424 | ompConstruct.u); |
425 | } |
426 | |
427 | /// Populate the global \see hostEvalInfo after processing clauses for the given |
428 | /// \p eval OpenMP target construct, or nested constructs, if these must be |
429 | /// evaluated outside of the target region per the spec. |
430 | /// |
431 | /// In particular, this will ensure that in 'target teams' and equivalent nested |
432 | /// constructs, the \c thread_limit and \c num_teams clauses will be evaluated |
433 | /// in the host. Additionally, loop bounds, steps and the \c num_threads clause |
434 | /// will also be evaluated in the host if a target SPMD construct is detected |
435 | /// (i.e. 'target teams distribute parallel do [simd]' or equivalent nesting). |
436 | /// |
437 | /// The result, stored as a global, is intended to be used to populate the \c |
438 | /// host_eval operands of the associated \c omp.target operation, and also to be |
439 | /// checked and used by later lowering steps to populate the corresponding |
440 | /// operands of the \c omp.teams, \c omp.parallel or \c omp.loop_nest |
441 | /// operations. |
442 | static void processHostEvalClauses(lower::AbstractConverter &converter, |
443 | semantics::SemanticsContext &semaCtx, |
444 | lower::StatementContext &stmtCtx, |
445 | lower::pft::Evaluation &eval, |
446 | mlir::Location loc) { |
447 | // Obtain the list of clauses of the given OpenMP block or loop construct |
448 | // evaluation. Other evaluations passed to this lambda keep `clauses` |
449 | // unchanged. |
450 | auto = [&semaCtx](lower::pft::Evaluation &eval, |
451 | List<Clause> &clauses) { |
452 | const auto *ompEval = eval.getIf<parser::OpenMPConstruct>(); |
453 | if (!ompEval) |
454 | return; |
455 | |
456 | const parser::OmpClauseList *beginClauseList = nullptr; |
457 | const parser::OmpClauseList *endClauseList = nullptr; |
458 | common::visit( |
459 | common::visitors{ |
460 | [&](const parser::OpenMPBlockConstruct &ompConstruct) { |
461 | const auto &beginDirective = |
462 | std::get<parser::OmpBeginBlockDirective>(ompConstruct.t); |
463 | beginClauseList = |
464 | &std::get<parser::OmpClauseList>(beginDirective.t); |
465 | endClauseList = &std::get<parser::OmpClauseList>( |
466 | std::get<parser::OmpEndBlockDirective>(ompConstruct.t).t); |
467 | }, |
468 | [&](const parser::OpenMPLoopConstruct &ompConstruct) { |
469 | const auto &beginDirective = |
470 | std::get<parser::OmpBeginLoopDirective>(ompConstruct.t); |
471 | beginClauseList = |
472 | &std::get<parser::OmpClauseList>(beginDirective.t); |
473 | |
474 | if (auto &endDirective = |
475 | std::get<std::optional<parser::OmpEndLoopDirective>>( |
476 | ompConstruct.t)) |
477 | endClauseList = |
478 | &std::get<parser::OmpClauseList>(endDirective->t); |
479 | }, |
480 | [&](const auto &) {}}, |
481 | ompEval->u); |
482 | |
483 | assert(beginClauseList && "expected begin directive" ); |
484 | clauses.append(makeClauses(*beginClauseList, semaCtx)); |
485 | |
486 | if (endClauseList) |
487 | clauses.append(makeClauses(*endClauseList, semaCtx)); |
488 | }; |
489 | |
490 | // Return the directive that is immediately nested inside of the given |
491 | // `parent` evaluation, if it is its only non-end-statement nested evaluation |
492 | // and it represents an OpenMP construct. |
493 | auto = [](lower::pft::Evaluation &parent) |
494 | -> std::optional<llvm::omp::Directive> { |
495 | if (!parent.hasNestedEvaluations()) |
496 | return std::nullopt; |
497 | |
498 | llvm::omp::Directive dir; |
499 | auto &nested = parent.getFirstNestedEvaluation(); |
500 | if (const auto *ompEval = nested.getIf<parser::OpenMPConstruct>()) |
501 | dir = extractOmpDirective(*ompEval); |
502 | else |
503 | return std::nullopt; |
504 | |
505 | for (auto &sibling : parent.getNestedEvaluations()) |
506 | if (&sibling != &nested && !sibling.isEndStmt()) |
507 | return std::nullopt; |
508 | |
509 | return dir; |
510 | }; |
511 | |
512 | // Process the given evaluation assuming it's part of a 'target' construct or |
513 | // captured by one, and store results in the global `hostEvalInfo`. |
514 | std::function<void(lower::pft::Evaluation &, const List<Clause> &)> |
515 | processEval; |
516 | processEval = [&](lower::pft::Evaluation &eval, const List<Clause> &clauses) { |
517 | using namespace llvm::omp; |
518 | ClauseProcessor cp(converter, semaCtx, clauses); |
519 | |
520 | // Call `processEval` recursively with the immediately nested evaluation and |
521 | // its corresponding clauses if there is a single nested evaluation |
522 | // representing an OpenMP directive that passes the given test. |
523 | auto processSingleNestedIf = [&](llvm::function_ref<bool(Directive)> test) { |
524 | std::optional<Directive> nestedDir = extractOnlyOmpNestedDir(eval); |
525 | if (!nestedDir || !test(*nestedDir)) |
526 | return; |
527 | |
528 | lower::pft::Evaluation &nestedEval = eval.getFirstNestedEvaluation(); |
529 | List<lower::omp::Clause> nestedClauses; |
530 | extractClauses(nestedEval, nestedClauses); |
531 | processEval(nestedEval, nestedClauses); |
532 | }; |
533 | |
534 | const auto *ompEval = eval.getIf<parser::OpenMPConstruct>(); |
535 | if (!ompEval) |
536 | return; |
537 | |
538 | HostEvalInfo &hostInfo = hostEvalInfo.back(); |
539 | |
540 | switch (extractOmpDirective(*ompEval)) { |
541 | case OMPD_teams_distribute_parallel_do: |
542 | case OMPD_teams_distribute_parallel_do_simd: |
543 | cp.processThreadLimit(stmtCtx, hostInfo.ops); |
544 | [[fallthrough]]; |
545 | case OMPD_target_teams_distribute_parallel_do: |
546 | case OMPD_target_teams_distribute_parallel_do_simd: |
547 | cp.processNumTeams(stmtCtx, hostInfo.ops); |
548 | [[fallthrough]]; |
549 | case OMPD_distribute_parallel_do: |
550 | case OMPD_distribute_parallel_do_simd: |
551 | cp.processNumThreads(stmtCtx, hostInfo.ops); |
552 | [[fallthrough]]; |
553 | case OMPD_distribute: |
554 | case OMPD_distribute_simd: |
555 | cp.processCollapse(loc, eval, hostInfo.ops, hostInfo.iv); |
556 | break; |
557 | |
558 | case OMPD_teams: |
559 | cp.processThreadLimit(stmtCtx, hostInfo.ops); |
560 | [[fallthrough]]; |
561 | case OMPD_target_teams: |
562 | cp.processNumTeams(stmtCtx, hostInfo.ops); |
563 | processSingleNestedIf([](Directive nestedDir) { |
564 | return topDistributeSet.test(nestedDir) || topLoopSet.test(nestedDir); |
565 | }); |
566 | break; |
567 | |
568 | case OMPD_teams_distribute: |
569 | case OMPD_teams_distribute_simd: |
570 | cp.processThreadLimit(stmtCtx, hostInfo.ops); |
571 | [[fallthrough]]; |
572 | case OMPD_target_teams_distribute: |
573 | case OMPD_target_teams_distribute_simd: |
574 | cp.processCollapse(loc, eval, hostInfo.ops, hostInfo.iv); |
575 | cp.processNumTeams(stmtCtx, hostInfo.ops); |
576 | break; |
577 | |
578 | case OMPD_teams_loop: |
579 | cp.processThreadLimit(stmtCtx, hostInfo.ops); |
580 | [[fallthrough]]; |
581 | case OMPD_target_teams_loop: |
582 | cp.processNumTeams(stmtCtx, hostInfo.ops); |
583 | [[fallthrough]]; |
584 | case OMPD_loop: |
585 | cp.processCollapse(loc, eval, hostInfo.ops, hostInfo.iv); |
586 | break; |
587 | |
588 | // Standalone 'target' case. |
589 | case OMPD_target: { |
590 | processSingleNestedIf( |
591 | [](Directive nestedDir) { return topTeamsSet.test(nestedDir); }); |
592 | break; |
593 | } |
594 | default: |
595 | break; |
596 | } |
597 | }; |
598 | |
599 | assert(!hostEvalInfo.empty() && "expected HOST_EVAL info structure" ); |
600 | |
601 | const auto *ompEval = eval.getIf<parser::OpenMPConstruct>(); |
602 | assert(ompEval && |
603 | llvm::omp::allTargetSet.test(extractOmpDirective(*ompEval)) && |
604 | "expected TARGET construct evaluation" ); |
605 | (void)ompEval; |
606 | |
607 | // Use the whole list of clauses passed to the construct here, rather than the |
608 | // ones only applied to omp.target. |
609 | List<lower::omp::Clause> clauses; |
610 | extractClauses(eval, clauses); |
611 | processEval(eval, clauses); |
612 | } |
613 | |
614 | static lower::pft::Evaluation * |
615 | getCollapsedLoopEval(lower::pft::Evaluation &eval, int collapseValue) { |
616 | // Return the Evaluation of the innermost collapsed loop, or the current one |
617 | // if there was no COLLAPSE. |
618 | if (collapseValue == 0) |
619 | return &eval; |
620 | |
621 | lower::pft::Evaluation *curEval = &eval.getFirstNestedEvaluation(); |
622 | for (int i = 1; i < collapseValue; i++) { |
623 | // The nested evaluations should be DoConstructs (i.e. they should form |
624 | // a loop nest). Each DoConstruct is a tuple <NonLabelDoStmt, Block, |
625 | // EndDoStmt>. |
626 | assert(curEval->isA<parser::DoConstruct>()); |
627 | curEval = &*std::next(curEval->getNestedEvaluations().begin()); |
628 | } |
629 | return curEval; |
630 | } |
631 | |
632 | static void genNestedEvaluations(lower::AbstractConverter &converter, |
633 | lower::pft::Evaluation &eval, |
634 | int collapseValue = 0) { |
635 | lower::pft::Evaluation *curEval = getCollapsedLoopEval(eval, collapseValue); |
636 | |
637 | for (lower::pft::Evaluation &e : curEval->getNestedEvaluations()) |
638 | converter.genEval(e); |
639 | } |
640 | |
641 | static fir::GlobalOp globalInitialization(lower::AbstractConverter &converter, |
642 | fir::FirOpBuilder &firOpBuilder, |
643 | const semantics::Symbol &sym, |
644 | const lower::pft::Variable &var, |
645 | mlir::Location currentLocation) { |
646 | std::string globalName = converter.mangleName(sym); |
647 | mlir::StringAttr linkage = firOpBuilder.createInternalLinkage(); |
648 | return Fortran::lower::defineGlobal(converter, var, globalName, linkage); |
649 | } |
650 | |
651 | // Get the extended value for \p val by extracting additional variable |
652 | // information from \p base. |
653 | static fir::ExtendedValue getExtendedValue(fir::ExtendedValue base, |
654 | mlir::Value val) { |
655 | return base.match( |
656 | [&](const fir::MutableBoxValue &box) -> fir::ExtendedValue { |
657 | return fir::MutableBoxValue(val, box.nonDeferredLenParams(), {}); |
658 | }, |
659 | [&](const auto &) -> fir::ExtendedValue { |
660 | return fir::substBase(base, val); |
661 | }); |
662 | } |
663 | |
664 | #ifndef NDEBUG |
665 | static bool isThreadPrivate(lower::SymbolRef sym) { |
666 | if (const auto *details = sym->detailsIf<semantics::CommonBlockDetails>()) { |
667 | for (const auto &obj : details->objects()) |
668 | if (!obj->test(semantics::Symbol::Flag::OmpThreadprivate)) |
669 | return false; |
670 | return true; |
671 | } |
672 | return sym->test(semantics::Symbol::Flag::OmpThreadprivate); |
673 | } |
674 | #endif |
675 | |
676 | static void threadPrivatizeVars(lower::AbstractConverter &converter, |
677 | lower::pft::Evaluation &eval) { |
678 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
679 | mlir::Location currentLocation = converter.getCurrentLocation(); |
680 | mlir::OpBuilder::InsertionGuard guard(firOpBuilder); |
681 | firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock()); |
682 | |
683 | // If the symbol corresponds to the original ThreadprivateOp, use the symbol |
684 | // value from that operation to create one ThreadprivateOp copy operation |
685 | // inside the parallel region. |
686 | // In some cases, however, the symbol will correspond to the original, |
687 | // non-threadprivate variable. This can happen, for instance, with a common |
688 | // block, declared in a separate module, used by a parent procedure and |
689 | // privatized in its child procedure. |
690 | auto genThreadprivateOp = [&](lower::SymbolRef sym) -> mlir::Value { |
691 | assert(isThreadPrivate(sym)); |
692 | mlir::Value symValue = converter.getSymbolAddress(sym); |
693 | mlir::Operation *op = symValue.getDefiningOp(); |
694 | if (auto declOp = mlir::dyn_cast<hlfir::DeclareOp>(op)) |
695 | op = declOp.getMemref().getDefiningOp(); |
696 | if (mlir::isa<mlir::omp::ThreadprivateOp>(op)) |
697 | symValue = mlir::dyn_cast<mlir::omp::ThreadprivateOp>(op).getSymAddr(); |
698 | return firOpBuilder.create<mlir::omp::ThreadprivateOp>( |
699 | currentLocation, symValue.getType(), symValue); |
700 | }; |
701 | |
702 | llvm::SetVector<const semantics::Symbol *> threadprivateSyms; |
703 | converter.collectSymbolSet(eval, threadprivateSyms, |
704 | semantics::Symbol::Flag::OmpThreadprivate, |
705 | /*collectSymbols=*/true, |
706 | /*collectHostAssociatedSymbols=*/true); |
707 | std::set<semantics::SourceName> threadprivateSymNames; |
708 | |
709 | // For a COMMON block, the ThreadprivateOp is generated for itself instead of |
710 | // its members, so only bind the value of the new copied ThreadprivateOp |
711 | // inside the parallel region to the common block symbol only once for |
712 | // multiple members in one COMMON block. |
713 | llvm::SetVector<const semantics::Symbol *> commonSyms; |
714 | for (std::size_t i = 0; i < threadprivateSyms.size(); i++) { |
715 | const semantics::Symbol *sym = threadprivateSyms[i]; |
716 | mlir::Value symThreadprivateValue; |
717 | // The variable may be used more than once, and each reference has one |
718 | // symbol with the same name. Only do once for references of one variable. |
719 | if (threadprivateSymNames.find(sym->name()) != threadprivateSymNames.end()) |
720 | continue; |
721 | threadprivateSymNames.insert(sym->name()); |
722 | if (const semantics::Symbol *common = |
723 | semantics::FindCommonBlockContaining(sym->GetUltimate())) { |
724 | mlir::Value commonThreadprivateValue; |
725 | if (commonSyms.contains(common)) { |
726 | commonThreadprivateValue = converter.getSymbolAddress(*common); |
727 | } else { |
728 | commonThreadprivateValue = genThreadprivateOp(*common); |
729 | converter.bindSymbol(*common, commonThreadprivateValue); |
730 | commonSyms.insert(common); |
731 | } |
732 | symThreadprivateValue = lower::genCommonBlockMember( |
733 | converter, currentLocation, sym->GetUltimate(), |
734 | commonThreadprivateValue); |
735 | } else { |
736 | symThreadprivateValue = genThreadprivateOp(*sym); |
737 | } |
738 | |
739 | fir::ExtendedValue sexv = converter.getSymbolExtendedValue(*sym); |
740 | fir::ExtendedValue symThreadprivateExv = |
741 | getExtendedValue(sexv, symThreadprivateValue); |
742 | converter.bindSymbol(*sym, symThreadprivateExv); |
743 | } |
744 | } |
745 | |
746 | static mlir::Operation * |
747 | createAndSetPrivatizedLoopVar(lower::AbstractConverter &converter, |
748 | mlir::Location loc, mlir::Value indexVal, |
749 | const semantics::Symbol *sym) { |
750 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
751 | mlir::OpBuilder::InsertPoint insPt = firOpBuilder.saveInsertionPoint(); |
752 | firOpBuilder.setInsertionPointToStart(firOpBuilder.getAllocaBlock()); |
753 | |
754 | mlir::Type tempTy = converter.genType(*sym); |
755 | |
756 | assert(converter.isPresentShallowLookup(*sym) && |
757 | "Expected symbol to be in symbol table." ); |
758 | |
759 | firOpBuilder.restoreInsertionPoint(insPt); |
760 | mlir::Value cvtVal = firOpBuilder.createConvert(loc, tempTy, indexVal); |
761 | hlfir::Entity lhs{converter.getSymbolAddress(*sym)}; |
762 | |
763 | lhs = hlfir::derefPointersAndAllocatables(loc, firOpBuilder, lhs); |
764 | |
765 | mlir::Operation *storeOp = |
766 | firOpBuilder.create<hlfir::AssignOp>(loc, cvtVal, lhs); |
767 | return storeOp; |
768 | } |
769 | |
770 | // This helper function implements the functionality of "promoting" non-CPTR |
771 | // arguments of use_device_ptr to use_device_addr arguments (automagic |
772 | // conversion of use_device_ptr -> use_device_addr in these cases). The way we |
773 | // do so currently is through the shuffling of operands from the |
774 | // devicePtrOperands to deviceAddrOperands, as well as the types, locations and |
775 | // symbols. |
776 | // |
777 | // This effectively implements some deprecated OpenMP functionality that some |
778 | // legacy applications unfortunately depend on (deprecated in specification |
779 | // version 5.2): |
780 | // |
781 | // "If a list item in a use_device_ptr clause is not of type C_PTR, the behavior |
782 | // is as if the list item appeared in a use_device_addr clause. Support for |
783 | // such list items in a use_device_ptr clause is deprecated." |
784 | static void promoteNonCPtrUseDevicePtrArgsToUseDeviceAddr( |
785 | llvm::SmallVectorImpl<mlir::Value> &useDeviceAddrVars, |
786 | llvm::SmallVectorImpl<const semantics::Symbol *> &useDeviceAddrSyms, |
787 | llvm::SmallVectorImpl<mlir::Value> &useDevicePtrVars, |
788 | llvm::SmallVectorImpl<const semantics::Symbol *> &useDevicePtrSyms) { |
789 | // Iterate over our use_device_ptr list and shift all non-cptr arguments into |
790 | // use_device_addr. |
791 | auto *varIt = useDevicePtrVars.begin(); |
792 | auto *symIt = useDevicePtrSyms.begin(); |
793 | while (varIt != useDevicePtrVars.end()) { |
794 | if (fir::isa_builtin_cptr_type(fir::unwrapRefType(varIt->getType()))) { |
795 | ++varIt; |
796 | ++symIt; |
797 | continue; |
798 | } |
799 | |
800 | useDeviceAddrVars.push_back(*varIt); |
801 | useDeviceAddrSyms.push_back(Elt: *symIt); |
802 | |
803 | varIt = useDevicePtrVars.erase(varIt); |
804 | symIt = useDevicePtrSyms.erase(CI: symIt); |
805 | } |
806 | } |
807 | |
808 | /// Extract the list of function and variable symbols affected by the given |
809 | /// 'declare target' directive and return the intended device type for them. |
810 | static void getDeclareTargetInfo( |
811 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
812 | lower::pft::Evaluation &eval, |
813 | const parser::OpenMPDeclareTargetConstruct &declareTargetConstruct, |
814 | mlir::omp::DeclareTargetOperands &clauseOps, |
815 | llvm::SmallVectorImpl<DeclareTargetCapturePair> &symbolAndClause) { |
816 | const auto &spec = |
817 | std::get<parser::OmpDeclareTargetSpecifier>(declareTargetConstruct.t); |
818 | if (const auto *objectList{parser::Unwrap<parser::OmpObjectList>(spec.u)}) { |
819 | ObjectList objects{makeObjects(*objectList, semaCtx)}; |
820 | // Case: declare target(func, var1, var2) |
821 | gatherFuncAndVarSyms(objects, mlir::omp::DeclareTargetCaptureClause::to, |
822 | symbolAndClause); |
823 | } else if (const auto *clauseList{ |
824 | parser::Unwrap<parser::OmpClauseList>(spec.u)}) { |
825 | List<Clause> clauses = makeClauses(*clauseList, semaCtx); |
826 | if (clauses.empty()) { |
827 | Fortran::lower::pft::FunctionLikeUnit *owningProc = |
828 | eval.getOwningProcedure(); |
829 | if (owningProc && (!owningProc->isMainProgram() || |
830 | owningProc->getMainProgramSymbol())) { |
831 | // Case: declare target, implicit capture of function |
832 | symbolAndClause.emplace_back(mlir::omp::DeclareTargetCaptureClause::to, |
833 | owningProc->getSubprogramSymbol()); |
834 | } |
835 | } |
836 | |
837 | ClauseProcessor cp(converter, semaCtx, clauses); |
838 | cp.processDeviceType(clauseOps); |
839 | cp.processEnter(symbolAndClause); |
840 | cp.processLink(symbolAndClause); |
841 | cp.processTo(symbolAndClause); |
842 | |
843 | cp.processTODO<clause::Indirect>(converter.getCurrentLocation(), |
844 | llvm::omp::Directive::OMPD_declare_target); |
845 | } |
846 | } |
847 | |
848 | static void collectDeferredDeclareTargets( |
849 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
850 | lower::pft::Evaluation &eval, |
851 | const parser::OpenMPDeclareTargetConstruct &declareTargetConstruct, |
852 | llvm::SmallVectorImpl<lower::OMPDeferredDeclareTargetInfo> |
853 | &deferredDeclareTarget) { |
854 | mlir::omp::DeclareTargetOperands clauseOps; |
855 | llvm::SmallVector<DeclareTargetCapturePair> symbolAndClause; |
856 | getDeclareTargetInfo(converter, semaCtx, eval, declareTargetConstruct, |
857 | clauseOps, symbolAndClause); |
858 | // Return the device type only if at least one of the targets for the |
859 | // directive is a function or subroutine |
860 | mlir::ModuleOp mod = converter.getFirOpBuilder().getModule(); |
861 | |
862 | for (const DeclareTargetCapturePair &symClause : symbolAndClause) { |
863 | mlir::Operation *op = mod.lookupSymbol( |
864 | converter.mangleName(std::get<const semantics::Symbol &>(symClause))); |
865 | |
866 | if (!op) { |
867 | deferredDeclareTarget.push_back({std::get<0>(symClause), |
868 | clauseOps.deviceType, |
869 | std::get<1>(symClause)}); |
870 | } |
871 | } |
872 | } |
873 | |
874 | static std::optional<mlir::omp::DeclareTargetDeviceType> |
875 | getDeclareTargetFunctionDevice( |
876 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
877 | lower::pft::Evaluation &eval, |
878 | const parser::OpenMPDeclareTargetConstruct &declareTargetConstruct) { |
879 | mlir::omp::DeclareTargetOperands clauseOps; |
880 | llvm::SmallVector<DeclareTargetCapturePair> symbolAndClause; |
881 | getDeclareTargetInfo(converter, semaCtx, eval, declareTargetConstruct, |
882 | clauseOps, symbolAndClause); |
883 | |
884 | // Return the device type only if at least one of the targets for the |
885 | // directive is a function or subroutine |
886 | mlir::ModuleOp mod = converter.getFirOpBuilder().getModule(); |
887 | for (const DeclareTargetCapturePair &symClause : symbolAndClause) { |
888 | mlir::Operation *op = mod.lookupSymbol( |
889 | converter.mangleName(std::get<const semantics::Symbol &>(symClause))); |
890 | |
891 | if (mlir::isa_and_nonnull<mlir::func::FuncOp>(op)) |
892 | return clauseOps.deviceType; |
893 | } |
894 | |
895 | return std::nullopt; |
896 | } |
897 | |
898 | /// Set up the entry block of the given `omp.loop_nest` operation, adding a |
899 | /// block argument for each loop induction variable and allocating and |
900 | /// initializing a private value to hold each of them. |
901 | /// |
902 | /// This function can also bind the symbols of any variables that should match |
903 | /// block arguments on parent loop wrapper operations attached to the same |
904 | /// loop. This allows the introduction of any necessary `hlfir.declare` |
905 | /// operations inside of the entry block of the `omp.loop_nest` operation and |
906 | /// not directly under any of the wrappers, which would invalidate them. |
907 | /// |
908 | /// \param [in] op - the loop nest operation. |
909 | /// \param [in] converter - PFT to MLIR conversion interface. |
910 | /// \param [in] loc - location. |
911 | /// \param [in] args - symbols of induction variables. |
912 | /// \param [in] wrapperArgs - list of parent loop wrappers and their associated |
913 | /// entry block arguments. |
914 | static void genLoopVars( |
915 | mlir::Operation *op, lower::AbstractConverter &converter, |
916 | mlir::Location &loc, llvm::ArrayRef<const semantics::Symbol *> args, |
917 | llvm::ArrayRef< |
918 | std::pair<mlir::omp::BlockArgOpenMPOpInterface, const EntryBlockArgs &>> |
919 | wrapperArgs = {}) { |
920 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
921 | auto ®ion = op->getRegion(0); |
922 | |
923 | std::size_t loopVarTypeSize = 0; |
924 | for (const semantics::Symbol *arg : args) |
925 | loopVarTypeSize = std::max(loopVarTypeSize, arg->GetUltimate().size()); |
926 | mlir::Type loopVarType = getLoopVarType(converter, loopVarTypeSize); |
927 | llvm::SmallVector<mlir::Type> tiv(args.size(), loopVarType); |
928 | llvm::SmallVector<mlir::Location> locs(args.size(), loc); |
929 | firOpBuilder.createBlock(®ion, {}, tiv, locs); |
930 | |
931 | // Update nested wrapper operands if parent wrappers have mapped these values |
932 | // to block arguments. |
933 | // |
934 | // Binding these values earlier would take care of this, but we cannot rely on |
935 | // that approach because binding in between the creation of a wrapper and the |
936 | // next one would result in 'hlfir.declare' operations being introduced inside |
937 | // of a wrapper, which is illegal. |
938 | mlir::IRMapping mapper; |
939 | for (auto [argGeneratingOp, blockArgs] : wrapperArgs) { |
940 | for (mlir::OpOperand &operand : argGeneratingOp->getOpOperands()) |
941 | operand.set(mapper.lookupOrDefault(operand.get())); |
942 | |
943 | for (const auto [arg, var] : llvm::zip_equal( |
944 | argGeneratingOp->getRegion(0).getArguments(), blockArgs.getVars())) |
945 | mapper.map(var, arg); |
946 | } |
947 | |
948 | // Bind the entry block arguments of parent wrappers to the corresponding |
949 | // symbols. |
950 | for (auto [argGeneratingOp, blockArgs] : wrapperArgs) |
951 | bindEntryBlockArgs(converter, argGeneratingOp, blockArgs); |
952 | |
953 | // The argument is not currently in memory, so make a temporary for the |
954 | // argument, and store it there, then bind that location to the argument. |
955 | mlir::Operation *storeOp = nullptr; |
956 | for (auto [argIndex, argSymbol] : llvm::enumerate(First&: args)) { |
957 | mlir::Value indexVal = fir::getBase(region.front().getArgument(argIndex)); |
958 | storeOp = |
959 | createAndSetPrivatizedLoopVar(converter, loc, indexVal, argSymbol); |
960 | } |
961 | firOpBuilder.setInsertionPointAfter(storeOp); |
962 | } |
963 | |
964 | static clause::Defaultmap::ImplicitBehavior |
965 | getDefaultmapIfPresent(const DefaultMapsTy &defaultMaps, mlir::Type varType) { |
966 | using DefMap = clause::Defaultmap; |
967 | |
968 | if (defaultMaps.empty()) |
969 | return DefMap::ImplicitBehavior::Default; |
970 | |
971 | if (llvm::is_contained(defaultMaps, DefMap::VariableCategory::All)) |
972 | return defaultMaps.at(DefMap::VariableCategory::All); |
973 | |
974 | // NOTE: Unsure if complex and/or vector falls into a scalar type |
975 | // or aggregate, but the current default implicit behaviour is to |
976 | // treat them as such (c_ptr has its own behaviour, so perhaps |
977 | // being lumped in as a scalar isn't the right thing). |
978 | if ((fir::isa_trivial(varType) || fir::isa_char(varType) || |
979 | fir::isa_builtin_cptr_type(varType)) && |
980 | llvm::is_contained(defaultMaps, DefMap::VariableCategory::Scalar)) |
981 | return defaultMaps.at(DefMap::VariableCategory::Scalar); |
982 | |
983 | if (fir::isPointerType(varType) && |
984 | llvm::is_contained(defaultMaps, DefMap::VariableCategory::Pointer)) |
985 | return defaultMaps.at(DefMap::VariableCategory::Pointer); |
986 | |
987 | if (fir::isAllocatableType(varType) && |
988 | llvm::is_contained(defaultMaps, DefMap::VariableCategory::Allocatable)) |
989 | return defaultMaps.at(DefMap::VariableCategory::Allocatable); |
990 | |
991 | if (fir::isa_aggregate(varType) && |
992 | llvm::is_contained(defaultMaps, DefMap::VariableCategory::Aggregate)) |
993 | return defaultMaps.at(DefMap::VariableCategory::Aggregate); |
994 | |
995 | return DefMap::ImplicitBehavior::Default; |
996 | } |
997 | |
998 | static std::pair<llvm::omp::OpenMPOffloadMappingFlags, |
999 | mlir::omp::VariableCaptureKind> |
1000 | getImplicitMapTypeAndKind(fir::FirOpBuilder &firOpBuilder, |
1001 | lower::AbstractConverter &converter, |
1002 | const DefaultMapsTy &defaultMaps, mlir::Type varType, |
1003 | mlir::Location loc, const semantics::Symbol &sym) { |
1004 | using DefMap = clause::Defaultmap; |
1005 | // Check if a value of type `type` can be passed to the kernel by value. |
1006 | // All kernel parameters are of pointer type, so if the value can be |
1007 | // represented inside of a pointer, then it can be passed by value. |
1008 | auto isLiteralType = [&](mlir::Type type) { |
1009 | const mlir::DataLayout &dl = firOpBuilder.getDataLayout(); |
1010 | mlir::Type ptrTy = |
1011 | mlir::LLVM::LLVMPointerType::get(&converter.getMLIRContext()); |
1012 | uint64_t ptrSize = dl.getTypeSize(ptrTy); |
1013 | uint64_t ptrAlign = dl.getTypePreferredAlignment(ptrTy); |
1014 | |
1015 | auto [size, align] = fir::getTypeSizeAndAlignmentOrCrash( |
1016 | loc, type, dl, converter.getKindMap()); |
1017 | return size <= ptrSize && align <= ptrAlign; |
1018 | }; |
1019 | |
1020 | llvm::omp::OpenMPOffloadMappingFlags mapFlag = |
1021 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT; |
1022 | |
1023 | auto implicitBehaviour = getDefaultmapIfPresent(defaultMaps, varType); |
1024 | if (implicitBehaviour == DefMap::ImplicitBehavior::Default) { |
1025 | mlir::omp::VariableCaptureKind captureKind = |
1026 | mlir::omp::VariableCaptureKind::ByRef; |
1027 | |
1028 | // If a variable is specified in declare target link and if device |
1029 | // type is not specified as `nohost`, it needs to be mapped tofrom |
1030 | mlir::ModuleOp mod = firOpBuilder.getModule(); |
1031 | mlir::Operation *op = mod.lookupSymbol(converter.mangleName(sym)); |
1032 | auto declareTargetOp = |
1033 | llvm::dyn_cast_if_present<mlir::omp::DeclareTargetInterface>(op); |
1034 | if (declareTargetOp && declareTargetOp.isDeclareTarget()) { |
1035 | if (declareTargetOp.getDeclareTargetCaptureClause() == |
1036 | mlir::omp::DeclareTargetCaptureClause::link && |
1037 | declareTargetOp.getDeclareTargetDeviceType() != |
1038 | mlir::omp::DeclareTargetDeviceType::nohost) { |
1039 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO; |
1040 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_FROM; |
1041 | } |
1042 | } else if (fir::isa_trivial(varType) || fir::isa_char(varType)) { |
1043 | // Scalars behave as if they were "firstprivate". |
1044 | // TODO: Handle objects that are shared/lastprivate or were listed |
1045 | // in an in_reduction clause. |
1046 | if (isLiteralType(varType)) { |
1047 | captureKind = mlir::omp::VariableCaptureKind::ByCopy; |
1048 | } else { |
1049 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO; |
1050 | } |
1051 | } else if (!fir::isa_builtin_cptr_type(varType)) { |
1052 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO; |
1053 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_FROM; |
1054 | } |
1055 | return std::make_pair(mapFlag, captureKind); |
1056 | } |
1057 | |
1058 | switch (implicitBehaviour) { |
1059 | case DefMap::ImplicitBehavior::Alloc: |
1060 | return std::make_pair(llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_NONE, |
1061 | mlir::omp::VariableCaptureKind::ByRef); |
1062 | break; |
1063 | case DefMap::ImplicitBehavior::Firstprivate: |
1064 | case DefMap::ImplicitBehavior::None: |
1065 | TODO(loc, "Firstprivate and None are currently unsupported defaultmap " |
1066 | "behaviour" ); |
1067 | break; |
1068 | case DefMap::ImplicitBehavior::From: |
1069 | return std::make_pair(mapFlag |= |
1070 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_FROM, |
1071 | mlir::omp::VariableCaptureKind::ByRef); |
1072 | break; |
1073 | case DefMap::ImplicitBehavior::Present: |
1074 | return std::make_pair(mapFlag |= |
1075 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_PRESENT, |
1076 | mlir::omp::VariableCaptureKind::ByRef); |
1077 | break; |
1078 | case DefMap::ImplicitBehavior::To: |
1079 | return std::make_pair(mapFlag |= |
1080 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO, |
1081 | (fir::isa_trivial(varType) || fir::isa_char(varType)) |
1082 | ? mlir::omp::VariableCaptureKind::ByCopy |
1083 | : mlir::omp::VariableCaptureKind::ByRef); |
1084 | break; |
1085 | case DefMap::ImplicitBehavior::Tofrom: |
1086 | return std::make_pair(mapFlag |= |
1087 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_FROM | |
1088 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO, |
1089 | mlir::omp::VariableCaptureKind::ByRef); |
1090 | break; |
1091 | case DefMap::ImplicitBehavior::Default: |
1092 | llvm_unreachable( |
1093 | "Implicit None Behaviour Should Have Been Handled Earlier" ); |
1094 | break; |
1095 | } |
1096 | |
1097 | return std::make_pair(mapFlag |= |
1098 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_FROM | |
1099 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO, |
1100 | mlir::omp::VariableCaptureKind::ByRef); |
1101 | } |
1102 | |
1103 | static void |
1104 | markDeclareTarget(mlir::Operation *op, lower::AbstractConverter &converter, |
1105 | mlir::omp::DeclareTargetCaptureClause captureClause, |
1106 | mlir::omp::DeclareTargetDeviceType deviceType) { |
1107 | // TODO: Add support for program local variables with declare target applied |
1108 | auto declareTargetOp = llvm::dyn_cast<mlir::omp::DeclareTargetInterface>(op); |
1109 | if (!declareTargetOp) |
1110 | fir::emitFatalError( |
1111 | converter.getCurrentLocation(), |
1112 | "Attempt to apply declare target on unsupported operation" ); |
1113 | |
1114 | // The function or global already has a declare target applied to it, very |
1115 | // likely through implicit capture (usage in another declare target |
1116 | // function/subroutine). It should be marked as any if it has been assigned |
1117 | // both host and nohost, else we skip, as there is no change |
1118 | if (declareTargetOp.isDeclareTarget()) { |
1119 | if (declareTargetOp.getDeclareTargetDeviceType() != deviceType) |
1120 | declareTargetOp.setDeclareTarget(mlir::omp::DeclareTargetDeviceType::any, |
1121 | captureClause); |
1122 | return; |
1123 | } |
1124 | |
1125 | declareTargetOp.setDeclareTarget(deviceType, captureClause); |
1126 | } |
1127 | |
1128 | static bool isPointerAssignment(const evaluate::Assignment &assign) { |
1129 | return common::visit( |
1130 | common::visitors{ |
1131 | [](const evaluate::Assignment::BoundsSpec &) { return true; }, |
1132 | [](const evaluate::Assignment::BoundsRemapping &) { return true; }, |
1133 | [](const auto &) { return false; }, |
1134 | }, |
1135 | assign.u); |
1136 | } |
1137 | |
1138 | //===----------------------------------------------------------------------===// |
1139 | // Op body generation helper structures and functions |
1140 | //===----------------------------------------------------------------------===// |
1141 | |
1142 | struct OpWithBodyGenInfo { |
1143 | /// A type for a code-gen callback function. This takes as argument the op for |
1144 | /// which the code is being generated and returns the arguments of the op's |
1145 | /// region. |
1146 | using GenOMPRegionEntryCBFn = |
1147 | std::function<llvm::SmallVector<const semantics::Symbol *>( |
1148 | mlir::Operation *)>; |
1149 | |
1150 | OpWithBodyGenInfo(lower::AbstractConverter &converter, |
1151 | lower::SymMap &symTable, |
1152 | semantics::SemanticsContext &semaCtx, mlir::Location loc, |
1153 | lower::pft::Evaluation &eval, llvm::omp::Directive dir) |
1154 | : converter(converter), symTable(symTable), semaCtx(semaCtx), loc(loc), |
1155 | eval(eval), dir(dir) {} |
1156 | |
1157 | OpWithBodyGenInfo &setClauses(const List<Clause> *value) { |
1158 | clauses = value; |
1159 | return *this; |
1160 | } |
1161 | |
1162 | OpWithBodyGenInfo &setDataSharingProcessor(DataSharingProcessor *value) { |
1163 | dsp = value; |
1164 | return *this; |
1165 | } |
1166 | |
1167 | OpWithBodyGenInfo &setEntryBlockArgs(const EntryBlockArgs *value) { |
1168 | blockArgs = value; |
1169 | return *this; |
1170 | } |
1171 | |
1172 | OpWithBodyGenInfo &setGenRegionEntryCb(GenOMPRegionEntryCBFn value) { |
1173 | genRegionEntryCB = value; |
1174 | return *this; |
1175 | } |
1176 | |
1177 | OpWithBodyGenInfo &setGenSkeletonOnly(bool value) { |
1178 | genSkeletonOnly = value; |
1179 | return *this; |
1180 | } |
1181 | |
1182 | /// [inout] converter to use for the clauses. |
1183 | lower::AbstractConverter &converter; |
1184 | /// [in] Symbol table |
1185 | lower::SymMap &symTable; |
1186 | /// [in] Semantics context |
1187 | semantics::SemanticsContext &semaCtx; |
1188 | /// [in] location in source code. |
1189 | mlir::Location loc; |
1190 | /// [in] current PFT node/evaluation. |
1191 | lower::pft::Evaluation &eval; |
1192 | /// [in] leaf directive for which to generate the op body. |
1193 | llvm::omp::Directive dir; |
1194 | /// [in] list of clauses to process. |
1195 | const List<Clause> *clauses = nullptr; |
1196 | /// [in] if provided, processes the construct's data-sharing attributes. |
1197 | DataSharingProcessor *dsp = nullptr; |
1198 | /// [in] if provided, it is used to create the op's region entry block. It is |
1199 | /// overriden when a \see genRegionEntryCB is provided. This is only valid for |
1200 | /// operations implementing the \see mlir::omp::BlockArgOpenMPOpInterface. |
1201 | const EntryBlockArgs *blockArgs = nullptr; |
1202 | /// [in] if provided, it overrides the default op's region entry block |
1203 | /// creation. |
1204 | GenOMPRegionEntryCBFn genRegionEntryCB = nullptr; |
1205 | /// [in] if set to `true`, skip generating nested evaluations and dispatching |
1206 | /// any further leaf constructs. |
1207 | bool genSkeletonOnly = false; |
1208 | }; |
1209 | |
1210 | /// Create the body (block) for an OpenMP Operation. |
1211 | /// |
1212 | /// \param [in] op - the operation the body belongs to. |
1213 | /// \param [in] info - options controlling code-gen for the construction. |
1214 | /// \param [in] queue - work queue with nested constructs. |
1215 | /// \param [in] item - item in the queue to generate body for. |
1216 | static void createBodyOfOp(mlir::Operation &op, const OpWithBodyGenInfo &info, |
1217 | const ConstructQueue &queue, |
1218 | ConstructQueue::const_iterator item) { |
1219 | fir::FirOpBuilder &firOpBuilder = info.converter.getFirOpBuilder(); |
1220 | |
1221 | auto insertMarker = [](fir::FirOpBuilder &builder) { |
1222 | mlir::Value undef = builder.create<fir::UndefOp>(builder.getUnknownLoc(), |
1223 | builder.getIndexType()); |
1224 | return undef.getDefiningOp(); |
1225 | }; |
1226 | |
1227 | // Create the entry block for the region and collect its arguments for use |
1228 | // within the region. The entry block will be created as follows: |
1229 | // - By default, it will be empty and have no arguments. |
1230 | // - Operations implementing the omp::BlockArgOpenMPOpInterface can set the |
1231 | // `info.blockArgs` pointer so that block arguments will be those |
1232 | // corresponding to entry block argument-generating clauses. Binding of |
1233 | // Fortran symbols to the new MLIR values is done automatically. |
1234 | // - If the `info.genRegionEntryCB` callback is set, it takes precedence and |
1235 | // allows callers to manually create the entry block with its intended |
1236 | // list of arguments and to bind these arguments to their corresponding |
1237 | // Fortran symbols. This is used for e.g. loop induction variables. |
1238 | auto regionArgs = [&]() -> llvm::SmallVector<const semantics::Symbol *> { |
1239 | if (info.genRegionEntryCB) |
1240 | return info.genRegionEntryCB(&op); |
1241 | |
1242 | if (info.blockArgs) { |
1243 | genEntryBlock(firOpBuilder, *info.blockArgs, op.getRegion(0)); |
1244 | bindEntryBlockArgs(info.converter, |
1245 | llvm::cast<mlir::omp::BlockArgOpenMPOpInterface>(op), |
1246 | *info.blockArgs); |
1247 | return llvm::to_vector(info.blockArgs->getSyms()); |
1248 | } |
1249 | |
1250 | firOpBuilder.createBlock(&op.getRegion(0)); |
1251 | return {}; |
1252 | }(); |
1253 | |
1254 | // Mark the earliest insertion point. |
1255 | mlir::Operation *marker = insertMarker(firOpBuilder); |
1256 | |
1257 | // If it is an unstructured region, create empty blocks for all evaluations. |
1258 | if (lower::omp::isLastItemInQueue(item, queue) && |
1259 | info.eval.lowerAsUnstructured()) { |
1260 | lower::createEmptyRegionBlocks<mlir::omp::TerminatorOp, mlir::omp::YieldOp>( |
1261 | firOpBuilder, info.eval.getNestedEvaluations()); |
1262 | } |
1263 | |
1264 | // Start with privatization, so that the lowering of the nested |
1265 | // code will use the right symbols. |
1266 | bool isLoop = llvm::omp::getDirectiveAssociation(info.dir) == |
1267 | llvm::omp::Association::Loop; |
1268 | bool privatize = info.clauses; |
1269 | |
1270 | firOpBuilder.setInsertionPoint(marker); |
1271 | std::optional<DataSharingProcessor> tempDsp; |
1272 | if (privatize && !info.dsp) { |
1273 | tempDsp.emplace(info.converter, info.semaCtx, *info.clauses, info.eval, |
1274 | Fortran::lower::omp::isLastItemInQueue(item, queue), |
1275 | /*useDelayedPrivatization=*/false, info.symTable); |
1276 | tempDsp->processStep1(); |
1277 | } |
1278 | |
1279 | if (info.dir == llvm::omp::Directive::OMPD_parallel) { |
1280 | threadPrivatizeVars(converter&: info.converter, eval&: info.eval); |
1281 | if (info.clauses) { |
1282 | firOpBuilder.setInsertionPoint(marker); |
1283 | ClauseProcessor(info.converter, info.semaCtx, *info.clauses) |
1284 | .processCopyin(); |
1285 | } |
1286 | } |
1287 | |
1288 | if (!info.genSkeletonOnly) { |
1289 | if (ConstructQueue::const_iterator next = std::next(x: item); |
1290 | next != queue.end()) { |
1291 | genOMPDispatch(info.converter, info.symTable, info.semaCtx, info.eval, |
1292 | info.loc, queue, next); |
1293 | } else { |
1294 | // genFIR(Evaluation&) tries to patch up unterminated blocks, causing |
1295 | // a lot of complications for our approach if the terminator generation |
1296 | // is delayed past this point. Insert a temporary terminator here, then |
1297 | // delete it. |
1298 | firOpBuilder.setInsertionPointToEnd(&op.getRegion(0).back()); |
1299 | auto *temp = lower::genOpenMPTerminator(firOpBuilder, &op, info.loc); |
1300 | firOpBuilder.setInsertionPointAfter(marker); |
1301 | genNestedEvaluations(converter&: info.converter, eval&: info.eval); |
1302 | temp->erase(); |
1303 | } |
1304 | } |
1305 | |
1306 | // Get or create a unique exiting block from the given region, or |
1307 | // return nullptr if there is no exiting block. |
1308 | auto getUniqueExit = [&](mlir::Region ®ion) -> mlir::Block * { |
1309 | // Find the blocks where the OMP terminator should go. In simple cases |
1310 | // it is the single block in the operation's region. When the region |
1311 | // is more complicated, especially with unstructured control flow, there |
1312 | // may be multiple blocks, and some of them may have non-OMP terminators |
1313 | // resulting from lowering of the code contained within the operation. |
1314 | // All the remaining blocks are potential exit points from the op's region. |
1315 | // |
1316 | // Explicit control flow cannot exit any OpenMP region (other than via |
1317 | // STOP), and that is enforced by semantic checks prior to lowering. STOP |
1318 | // statements are lowered to a function call. |
1319 | |
1320 | // Collect unterminated blocks. |
1321 | llvm::SmallVector<mlir::Block *> exits; |
1322 | for (mlir::Block &b : region) { |
1323 | if (b.empty() || !b.back().hasTrait<mlir::OpTrait::IsTerminator>()) |
1324 | exits.push_back(&b); |
1325 | } |
1326 | |
1327 | if (exits.empty()) |
1328 | return nullptr; |
1329 | // If there already is a unique exiting block, do not create another one. |
1330 | // Additionally, some ops (e.g. omp.sections) require only 1 block in |
1331 | // its region. |
1332 | if (exits.size() == 1) |
1333 | return exits[0]; |
1334 | mlir::Block *exit = firOpBuilder.createBlock(®ion); |
1335 | for (mlir::Block *b : exits) { |
1336 | firOpBuilder.setInsertionPointToEnd(b); |
1337 | firOpBuilder.create<mlir::cf::BranchOp>(info.loc, exit); |
1338 | } |
1339 | return exit; |
1340 | }; |
1341 | |
1342 | if (auto *exitBlock = getUniqueExit(op.getRegion(0))) { |
1343 | firOpBuilder.setInsertionPointToEnd(exitBlock); |
1344 | auto *term = lower::genOpenMPTerminator(firOpBuilder, &op, info.loc); |
1345 | // Only insert lastprivate code when there actually is an exit block. |
1346 | // Such a block may not exist if the nested code produced an infinite |
1347 | // loop (this may not make sense in production code, but a user could |
1348 | // write that and we should handle it). |
1349 | firOpBuilder.setInsertionPoint(term); |
1350 | if (privatize) { |
1351 | // DataSharingProcessor::processStep2() may create operations before/after |
1352 | // the one passed as argument. We need to treat loop wrappers and their |
1353 | // nested loop as a unit, so we need to pass the bottom level wrapper (if |
1354 | // present). Otherwise, these operations will be inserted within a |
1355 | // wrapper region. |
1356 | mlir::Operation *privatizationBottomLevelOp = &op; |
1357 | if (auto loopNest = llvm::dyn_cast<mlir::omp::LoopNestOp>(op)) { |
1358 | llvm::SmallVector<mlir::omp::LoopWrapperInterface> wrappers; |
1359 | loopNest.gatherWrappers(wrappers); |
1360 | if (!wrappers.empty()) |
1361 | privatizationBottomLevelOp = &*wrappers.front(); |
1362 | } |
1363 | |
1364 | if (!info.dsp) { |
1365 | assert(tempDsp.has_value()); |
1366 | tempDsp->processStep2(privatizationBottomLevelOp, isLoop); |
1367 | } else { |
1368 | if (isLoop && regionArgs.size() > 0) { |
1369 | for (const auto ®ionArg : regionArgs) { |
1370 | info.dsp->pushLoopIV(info.converter.getSymbolAddress(*regionArg)); |
1371 | } |
1372 | } |
1373 | info.dsp->processStep2(privatizationBottomLevelOp, isLoop); |
1374 | } |
1375 | } |
1376 | } |
1377 | |
1378 | firOpBuilder.setInsertionPointAfter(marker); |
1379 | marker->erase(); |
1380 | } |
1381 | |
1382 | static void genBodyOfTargetDataOp( |
1383 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
1384 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
1385 | mlir::omp::TargetDataOp &dataOp, const EntryBlockArgs &args, |
1386 | const mlir::Location ¤tLocation, const ConstructQueue &queue, |
1387 | ConstructQueue::const_iterator item) { |
1388 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
1389 | |
1390 | genEntryBlock(firOpBuilder, args, dataOp.getRegion()); |
1391 | bindEntryBlockArgs(converter, dataOp, args); |
1392 | |
1393 | // Insert dummy instruction to remember the insertion position. The |
1394 | // marker will be deleted by clean up passes since there are no uses. |
1395 | // Remembering the position for further insertion is important since |
1396 | // there are hlfir.declares inserted above while setting block arguments |
1397 | // and new code from the body should be inserted after that. |
1398 | mlir::Value undefMarker = firOpBuilder.create<fir::UndefOp>( |
1399 | dataOp.getLoc(), firOpBuilder.getIndexType()); |
1400 | |
1401 | // Create blocks for unstructured regions. This has to be done since |
1402 | // blocks are initially allocated with the function as the parent region. |
1403 | if (eval.lowerAsUnstructured()) { |
1404 | lower::createEmptyRegionBlocks<mlir::omp::TerminatorOp, mlir::omp::YieldOp>( |
1405 | firOpBuilder, eval.getNestedEvaluations()); |
1406 | } |
1407 | |
1408 | firOpBuilder.create<mlir::omp::TerminatorOp>(currentLocation); |
1409 | |
1410 | // Set the insertion point after the marker. |
1411 | firOpBuilder.setInsertionPointAfter(undefMarker.getDefiningOp()); |
1412 | |
1413 | if (ConstructQueue::const_iterator next = std::next(x: item); |
1414 | next != queue.end()) { |
1415 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
1416 | next); |
1417 | } else { |
1418 | genNestedEvaluations(converter, eval); |
1419 | } |
1420 | } |
1421 | |
1422 | // This generates intermediate common block member accesses within a region |
1423 | // and then rebinds the members symbol to the intermediate accessors we have |
1424 | // generated so that subsequent code generation will utilise these instead. |
1425 | // |
1426 | // When the scope changes, the bindings to the intermediate accessors should |
1427 | // be dropped in place of the original symbol bindings. |
1428 | // |
1429 | // This is for utilisation with TargetOp. |
1430 | static void genIntermediateCommonBlockAccessors( |
1431 | Fortran::lower::AbstractConverter &converter, |
1432 | const mlir::Location ¤tLocation, |
1433 | llvm::ArrayRef<mlir::BlockArgument> mapBlockArgs, |
1434 | llvm::ArrayRef<const Fortran::semantics::Symbol *> mapSyms) { |
1435 | // Iterate over the symbol list, which will be shorter than the list of |
1436 | // arguments if new entry block arguments were introduced to implicitly map |
1437 | // outside values used by the bounds cloned into the target region. In that |
1438 | // case, the additional block arguments do not need processing here. |
1439 | for (auto [mapSym, mapArg] : llvm::zip_first(mapSyms, mapBlockArgs)) { |
1440 | auto *details = mapSym->detailsIf<Fortran::semantics::CommonBlockDetails>(); |
1441 | if (!details) |
1442 | continue; |
1443 | |
1444 | for (auto obj : details->objects()) { |
1445 | auto targetCBMemberBind = Fortran::lower::genCommonBlockMember( |
1446 | converter, currentLocation, *obj, mapArg); |
1447 | fir::ExtendedValue sexv = converter.getSymbolExtendedValue(*obj); |
1448 | fir::ExtendedValue targetCBExv = |
1449 | getExtendedValue(sexv, targetCBMemberBind); |
1450 | converter.bindSymbol(*obj, targetCBExv); |
1451 | } |
1452 | } |
1453 | } |
1454 | |
1455 | // This functions creates a block for the body of the targetOp's region. It adds |
1456 | // all the symbols present in mapSymbols as block arguments to this block. |
1457 | static void genBodyOfTargetOp( |
1458 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
1459 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
1460 | mlir::omp::TargetOp &targetOp, const EntryBlockArgs &args, |
1461 | const mlir::Location ¤tLocation, const ConstructQueue &queue, |
1462 | ConstructQueue::const_iterator item, DataSharingProcessor &dsp) { |
1463 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
1464 | auto argIface = llvm::cast<mlir::omp::BlockArgOpenMPOpInterface>(*targetOp); |
1465 | |
1466 | mlir::Region ®ion = targetOp.getRegion(); |
1467 | mlir::Block *entryBlock = genEntryBlock(firOpBuilder, args, region); |
1468 | bindEntryBlockArgs(converter, targetOp, args); |
1469 | if (!hostEvalInfo.empty()) |
1470 | hostEvalInfo.back().bindOperands(argIface.getHostEvalBlockArgs()); |
1471 | |
1472 | // Check if cloning the bounds introduced any dependency on the outer region. |
1473 | // If so, then either clone them as well if they are MemoryEffectFree, or else |
1474 | // copy them to a new temporary and add them to the map and block_argument |
1475 | // lists and replace their uses with the new temporary. |
1476 | llvm::SetVector<mlir::Value> valuesDefinedAbove; |
1477 | mlir::getUsedValuesDefinedAbove(region, valuesDefinedAbove); |
1478 | while (!valuesDefinedAbove.empty()) { |
1479 | for (mlir::Value val : valuesDefinedAbove) { |
1480 | mlir::Operation *valOp = val.getDefiningOp(); |
1481 | |
1482 | // NOTE: We skip BoxDimsOp's as the lesser of two evils is to map the |
1483 | // indices separately, as the alternative is to eventually map the Box, |
1484 | // which comes with a fairly large overhead comparatively. We could be |
1485 | // more robust about this and check using a BackwardsSlice to see if we |
1486 | // run the risk of mapping a box. |
1487 | if (valOp && mlir::isMemoryEffectFree(valOp) && |
1488 | !mlir::isa<fir::BoxDimsOp>(valOp)) { |
1489 | mlir::Operation *clonedOp = valOp->clone(); |
1490 | entryBlock->push_front(clonedOp); |
1491 | |
1492 | auto replace = [entryBlock](mlir::OpOperand &use) { |
1493 | return use.getOwner()->getBlock() == entryBlock; |
1494 | }; |
1495 | |
1496 | valOp->getResults().replaceUsesWithIf(clonedOp->getResults(), replace); |
1497 | valOp->replaceUsesWithIf(clonedOp, replace); |
1498 | } else { |
1499 | auto savedIP = firOpBuilder.getInsertionPoint(); |
1500 | |
1501 | if (valOp) |
1502 | firOpBuilder.setInsertionPointAfter(valOp); |
1503 | else |
1504 | // This means val is a block argument |
1505 | firOpBuilder.setInsertionPoint(targetOp); |
1506 | |
1507 | auto copyVal = |
1508 | firOpBuilder.createTemporary(val.getLoc(), val.getType()); |
1509 | firOpBuilder.createStoreWithConvert(copyVal.getLoc(), val, copyVal); |
1510 | |
1511 | fir::factory::AddrAndBoundsInfo info = |
1512 | fir::factory::getDataOperandBaseAddr( |
1513 | firOpBuilder, val, /*isOptional=*/false, val.getLoc()); |
1514 | llvm::SmallVector<mlir::Value> bounds = |
1515 | fir::factory::genImplicitBoundsOps<mlir::omp::MapBoundsOp, |
1516 | mlir::omp::MapBoundsType>( |
1517 | firOpBuilder, info, |
1518 | hlfir::translateToExtendedValue(val.getLoc(), firOpBuilder, |
1519 | hlfir::Entity{val}) |
1520 | .first, |
1521 | /*dataExvIsAssumedSize=*/false, val.getLoc()); |
1522 | |
1523 | std::stringstream name; |
1524 | firOpBuilder.setInsertionPoint(targetOp); |
1525 | |
1526 | llvm::omp::OpenMPOffloadMappingFlags mapFlag = |
1527 | llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_IMPLICIT; |
1528 | mlir::omp::VariableCaptureKind captureKind = |
1529 | mlir::omp::VariableCaptureKind::ByRef; |
1530 | |
1531 | mlir::Type eleType = copyVal.getType(); |
1532 | if (auto refType = |
1533 | mlir::dyn_cast<fir::ReferenceType>(copyVal.getType())) |
1534 | eleType = refType.getElementType(); |
1535 | |
1536 | if (fir::isa_trivial(eleType) || fir::isa_char(eleType)) { |
1537 | captureKind = mlir::omp::VariableCaptureKind::ByCopy; |
1538 | } else if (!fir::isa_builtin_cptr_type(eleType)) { |
1539 | mapFlag |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TO; |
1540 | } |
1541 | |
1542 | mlir::Value mapOp = createMapInfoOp( |
1543 | firOpBuilder, copyVal.getLoc(), copyVal, |
1544 | /*varPtrPtr=*/mlir::Value{}, name.str(), bounds, |
1545 | /*members=*/llvm::SmallVector<mlir::Value>{}, |
1546 | /*membersIndex=*/mlir::ArrayAttr{}, |
1547 | static_cast< |
1548 | std::underlying_type_t<llvm::omp::OpenMPOffloadMappingFlags>>( |
1549 | mapFlag), |
1550 | captureKind, copyVal.getType()); |
1551 | |
1552 | // Get the index of the first non-map argument before modifying mapVars, |
1553 | // then append an element to mapVars and an associated entry block |
1554 | // argument at that index. |
1555 | unsigned insertIndex = |
1556 | argIface.getMapBlockArgsStart() + argIface.numMapBlockArgs(); |
1557 | targetOp.getMapVarsMutable().append(mapOp); |
1558 | mlir::Value clonedValArg = region.insertArgument( |
1559 | insertIndex, copyVal.getType(), copyVal.getLoc()); |
1560 | |
1561 | firOpBuilder.setInsertionPointToStart(entryBlock); |
1562 | auto loadOp = firOpBuilder.create<fir::LoadOp>(clonedValArg.getLoc(), |
1563 | clonedValArg); |
1564 | val.replaceUsesWithIf(loadOp->getResult(0), |
1565 | [entryBlock](mlir::OpOperand &use) { |
1566 | return use.getOwner()->getBlock() == entryBlock; |
1567 | }); |
1568 | firOpBuilder.setInsertionPoint(entryBlock, savedIP); |
1569 | } |
1570 | } |
1571 | valuesDefinedAbove.clear(); |
1572 | mlir::getUsedValuesDefinedAbove(region, valuesDefinedAbove); |
1573 | } |
1574 | |
1575 | // Insert dummy instruction to remember the insertion position. The |
1576 | // marker will be deleted since there are not uses. |
1577 | // In the HLFIR flow there are hlfir.declares inserted above while |
1578 | // setting block arguments. |
1579 | mlir::Value undefMarker = firOpBuilder.create<fir::UndefOp>( |
1580 | targetOp.getLoc(), firOpBuilder.getIndexType()); |
1581 | |
1582 | // Create blocks for unstructured regions. This has to be done since |
1583 | // blocks are initially allocated with the function as the parent region. |
1584 | if (lower::omp::isLastItemInQueue(item, queue) && |
1585 | eval.lowerAsUnstructured()) { |
1586 | lower::createEmptyRegionBlocks<mlir::omp::TerminatorOp, mlir::omp::YieldOp>( |
1587 | firOpBuilder, eval.getNestedEvaluations()); |
1588 | } |
1589 | |
1590 | firOpBuilder.create<mlir::omp::TerminatorOp>(currentLocation); |
1591 | |
1592 | // Create the insertion point after the marker. |
1593 | firOpBuilder.setInsertionPointAfter(undefMarker.getDefiningOp()); |
1594 | |
1595 | // If we map a common block using it's symbol e.g. map(tofrom: /common_block/) |
1596 | // and accessing its members within the target region, there is a large |
1597 | // chance we will end up with uses external to the region accessing the common |
1598 | // resolve these, we do so by generating new common block member accesses |
1599 | // within the region, binding them to the member symbol for the scope of the |
1600 | // region so that subsequent code generation within the region will utilise |
1601 | // our new member accesses we have created. |
1602 | genIntermediateCommonBlockAccessors( |
1603 | converter, currentLocation, argIface.getMapBlockArgs(), args.map.syms); |
1604 | |
1605 | if (ConstructQueue::const_iterator next = std::next(x: item); |
1606 | next != queue.end()) { |
1607 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
1608 | next); |
1609 | } else { |
1610 | genNestedEvaluations(converter, eval); |
1611 | } |
1612 | |
1613 | dsp.processStep2(targetOp, /*isLoop=*/false); |
1614 | } |
1615 | |
1616 | template <typename OpTy, typename... Args> |
1617 | static OpTy genOpWithBody(const OpWithBodyGenInfo &info, |
1618 | const ConstructQueue &queue, |
1619 | ConstructQueue::const_iterator item, Args &&...args) { |
1620 | auto op = info.converter.getFirOpBuilder().create<OpTy>( |
1621 | info.loc, std::forward<Args>(args)...); |
1622 | createBodyOfOp(*op, info, queue, item); |
1623 | return op; |
1624 | } |
1625 | |
1626 | template <typename OpTy, typename ClauseOpsTy> |
1627 | static OpTy genWrapperOp(lower::AbstractConverter &converter, |
1628 | mlir::Location loc, const ClauseOpsTy &clauseOps, |
1629 | const EntryBlockArgs &args) { |
1630 | static_assert( |
1631 | OpTy::template hasTrait<mlir::omp::LoopWrapperInterface::Trait>(), |
1632 | "expected a loop wrapper" ); |
1633 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
1634 | |
1635 | // Create wrapper. |
1636 | auto op = firOpBuilder.create<OpTy>(loc, clauseOps); |
1637 | |
1638 | // Create entry block with arguments. |
1639 | genEntryBlock(firOpBuilder, args, op.getRegion()); |
1640 | |
1641 | return op; |
1642 | } |
1643 | |
1644 | //===----------------------------------------------------------------------===// |
1645 | // Code generation functions for clauses |
1646 | //===----------------------------------------------------------------------===// |
1647 | |
1648 | static void genCancelClauses(lower::AbstractConverter &converter, |
1649 | semantics::SemanticsContext &semaCtx, |
1650 | const List<Clause> &clauses, mlir::Location loc, |
1651 | mlir::omp::CancelOperands &clauseOps) { |
1652 | ClauseProcessor cp(converter, semaCtx, clauses); |
1653 | cp.processCancelDirectiveName(clauseOps); |
1654 | cp.processIf(llvm::omp::Directive::OMPD_cancel, clauseOps); |
1655 | } |
1656 | |
1657 | static void |
1658 | genCancellationPointClauses(lower::AbstractConverter &converter, |
1659 | semantics::SemanticsContext &semaCtx, |
1660 | const List<Clause> &clauses, mlir::Location loc, |
1661 | mlir::omp::CancellationPointOperands &clauseOps) { |
1662 | ClauseProcessor cp(converter, semaCtx, clauses); |
1663 | cp.processCancelDirectiveName(clauseOps); |
1664 | } |
1665 | |
1666 | static void genCriticalDeclareClauses( |
1667 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1668 | const List<Clause> &clauses, mlir::Location loc, |
1669 | mlir::omp::CriticalDeclareOperands &clauseOps, llvm::StringRef name) { |
1670 | ClauseProcessor cp(converter, semaCtx, clauses); |
1671 | cp.processHint(clauseOps); |
1672 | clauseOps.symName = |
1673 | mlir::StringAttr::get(converter.getFirOpBuilder().getContext(), name); |
1674 | } |
1675 | |
1676 | static void genDistributeClauses(lower::AbstractConverter &converter, |
1677 | semantics::SemanticsContext &semaCtx, |
1678 | lower::StatementContext &stmtCtx, |
1679 | const List<Clause> &clauses, |
1680 | mlir::Location loc, |
1681 | mlir::omp::DistributeOperands &clauseOps) { |
1682 | ClauseProcessor cp(converter, semaCtx, clauses); |
1683 | cp.processAllocate(clauseOps); |
1684 | cp.processDistSchedule(stmtCtx, clauseOps); |
1685 | cp.processOrder(clauseOps); |
1686 | } |
1687 | |
1688 | static void genFlushClauses(lower::AbstractConverter &converter, |
1689 | semantics::SemanticsContext &semaCtx, |
1690 | const ObjectList &objects, |
1691 | const List<Clause> &clauses, mlir::Location loc, |
1692 | llvm::SmallVectorImpl<mlir::Value> &operandRange) { |
1693 | if (!objects.empty()) |
1694 | genObjectList(objects, converter, operandRange); |
1695 | |
1696 | ClauseProcessor cp(converter, semaCtx, clauses); |
1697 | cp.processTODO<clause::AcqRel, clause::Acquire, clause::Release, |
1698 | clause::SeqCst>(loc, llvm::omp::OMPD_flush); |
1699 | } |
1700 | |
1701 | static void |
1702 | genLoopNestClauses(lower::AbstractConverter &converter, |
1703 | semantics::SemanticsContext &semaCtx, |
1704 | lower::pft::Evaluation &eval, const List<Clause> &clauses, |
1705 | mlir::Location loc, mlir::omp::LoopNestOperands &clauseOps, |
1706 | llvm::SmallVectorImpl<const semantics::Symbol *> &iv) { |
1707 | ClauseProcessor cp(converter, semaCtx, clauses); |
1708 | |
1709 | if (hostEvalInfo.empty() || !hostEvalInfo.back().apply(clauseOps, iv)) |
1710 | cp.processCollapse(loc, eval, clauseOps, iv); |
1711 | |
1712 | clauseOps.loopInclusive = converter.getFirOpBuilder().getUnitAttr(); |
1713 | } |
1714 | |
1715 | static void genLoopClauses( |
1716 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1717 | const List<Clause> &clauses, mlir::Location loc, |
1718 | mlir::omp::LoopOperands &clauseOps, |
1719 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1720 | ClauseProcessor cp(converter, semaCtx, clauses); |
1721 | cp.processBind(clauseOps); |
1722 | cp.processOrder(clauseOps); |
1723 | cp.processReduction(loc, clauseOps, reductionSyms); |
1724 | cp.processTODO<clause::Lastprivate>(loc, llvm::omp::Directive::OMPD_loop); |
1725 | } |
1726 | |
1727 | static void genMaskedClauses(lower::AbstractConverter &converter, |
1728 | semantics::SemanticsContext &semaCtx, |
1729 | lower::StatementContext &stmtCtx, |
1730 | const List<Clause> &clauses, mlir::Location loc, |
1731 | mlir::omp::MaskedOperands &clauseOps) { |
1732 | ClauseProcessor cp(converter, semaCtx, clauses); |
1733 | cp.processFilter(stmtCtx, clauseOps); |
1734 | } |
1735 | |
1736 | static void |
1737 | genOrderedRegionClauses(lower::AbstractConverter &converter, |
1738 | semantics::SemanticsContext &semaCtx, |
1739 | const List<Clause> &clauses, mlir::Location loc, |
1740 | mlir::omp::OrderedRegionOperands &clauseOps) { |
1741 | ClauseProcessor cp(converter, semaCtx, clauses); |
1742 | cp.processTODO<clause::Simd>(loc, llvm::omp::Directive::OMPD_ordered); |
1743 | } |
1744 | |
1745 | static void genParallelClauses( |
1746 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1747 | lower::StatementContext &stmtCtx, const List<Clause> &clauses, |
1748 | mlir::Location loc, mlir::omp::ParallelOperands &clauseOps, |
1749 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1750 | ClauseProcessor cp(converter, semaCtx, clauses); |
1751 | cp.processAllocate(clauseOps); |
1752 | cp.processIf(llvm::omp::Directive::OMPD_parallel, clauseOps); |
1753 | |
1754 | if (hostEvalInfo.empty() || !hostEvalInfo.back().apply(clauseOps)) |
1755 | cp.processNumThreads(stmtCtx, clauseOps); |
1756 | |
1757 | cp.processProcBind(clauseOps); |
1758 | cp.processReduction(loc, clauseOps, reductionSyms); |
1759 | } |
1760 | |
1761 | static void genScanClauses(lower::AbstractConverter &converter, |
1762 | semantics::SemanticsContext &semaCtx, |
1763 | const List<Clause> &clauses, mlir::Location loc, |
1764 | mlir::omp::ScanOperands &clauseOps) { |
1765 | ClauseProcessor cp(converter, semaCtx, clauses); |
1766 | cp.processInclusive(loc, clauseOps); |
1767 | cp.processExclusive(loc, clauseOps); |
1768 | } |
1769 | |
1770 | static void genSectionsClauses( |
1771 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1772 | const List<Clause> &clauses, mlir::Location loc, |
1773 | mlir::omp::SectionsOperands &clauseOps, |
1774 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1775 | ClauseProcessor cp(converter, semaCtx, clauses); |
1776 | cp.processAllocate(clauseOps); |
1777 | cp.processNowait(clauseOps); |
1778 | cp.processReduction(loc, clauseOps, reductionSyms); |
1779 | // TODO Support delayed privatization. |
1780 | } |
1781 | |
1782 | static void genSimdClauses( |
1783 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1784 | const List<Clause> &clauses, mlir::Location loc, |
1785 | mlir::omp::SimdOperands &clauseOps, |
1786 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1787 | ClauseProcessor cp(converter, semaCtx, clauses); |
1788 | cp.processAligned(clauseOps); |
1789 | cp.processIf(llvm::omp::Directive::OMPD_simd, clauseOps); |
1790 | cp.processNontemporal(clauseOps); |
1791 | cp.processOrder(clauseOps); |
1792 | cp.processReduction(loc, clauseOps, reductionSyms); |
1793 | cp.processSafelen(clauseOps); |
1794 | cp.processSimdlen(clauseOps); |
1795 | |
1796 | cp.processTODO<clause::Linear>(loc, llvm::omp::Directive::OMPD_simd); |
1797 | } |
1798 | |
1799 | static void genSingleClauses(lower::AbstractConverter &converter, |
1800 | semantics::SemanticsContext &semaCtx, |
1801 | const List<Clause> &clauses, mlir::Location loc, |
1802 | mlir::omp::SingleOperands &clauseOps) { |
1803 | ClauseProcessor cp(converter, semaCtx, clauses); |
1804 | cp.processAllocate(clauseOps); |
1805 | cp.processCopyprivate(loc, clauseOps); |
1806 | cp.processNowait(clauseOps); |
1807 | // TODO Support delayed privatization. |
1808 | } |
1809 | |
1810 | static void genTargetClauses( |
1811 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1812 | lower::SymMap &symTable, lower::StatementContext &stmtCtx, |
1813 | lower::pft::Evaluation &eval, const List<Clause> &clauses, |
1814 | mlir::Location loc, mlir::omp::TargetOperands &clauseOps, |
1815 | DefaultMapsTy &defaultMaps, |
1816 | llvm::SmallVectorImpl<const semantics::Symbol *> &hasDeviceAddrSyms, |
1817 | llvm::SmallVectorImpl<const semantics::Symbol *> &isDevicePtrSyms, |
1818 | llvm::SmallVectorImpl<const semantics::Symbol *> &mapSyms) { |
1819 | ClauseProcessor cp(converter, semaCtx, clauses); |
1820 | cp.processBare(clauseOps); |
1821 | cp.processDefaultMap(stmtCtx, defaultMaps); |
1822 | cp.processDepend(symTable, stmtCtx, clauseOps); |
1823 | cp.processDevice(stmtCtx, clauseOps); |
1824 | cp.processHasDeviceAddr(stmtCtx, clauseOps, hasDeviceAddrSyms); |
1825 | if (!hostEvalInfo.empty()) { |
1826 | // Only process host_eval if compiling for the host device. |
1827 | processHostEvalClauses(converter, semaCtx, stmtCtx, eval, loc); |
1828 | hostEvalInfo.back().collectValues(clauseOps.hostEvalVars); |
1829 | } |
1830 | cp.processIf(llvm::omp::Directive::OMPD_target, clauseOps); |
1831 | cp.processIsDevicePtr(clauseOps, isDevicePtrSyms); |
1832 | cp.processMap(loc, stmtCtx, clauseOps, &mapSyms); |
1833 | cp.processNowait(clauseOps); |
1834 | cp.processThreadLimit(stmtCtx, clauseOps); |
1835 | |
1836 | cp.processTODO<clause::Allocate, clause::InReduction, clause::UsesAllocators>( |
1837 | loc, llvm::omp::Directive::OMPD_target); |
1838 | |
1839 | // `target private(..)` is only supported in delayed privatization mode. |
1840 | if (!enableDelayedPrivatizationStaging) |
1841 | cp.processTODO<clause::Firstprivate, clause::Private>( |
1842 | loc, llvm::omp::Directive::OMPD_target); |
1843 | } |
1844 | |
1845 | static void genTargetDataClauses( |
1846 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1847 | lower::StatementContext &stmtCtx, const List<Clause> &clauses, |
1848 | mlir::Location loc, mlir::omp::TargetDataOperands &clauseOps, |
1849 | llvm::SmallVectorImpl<const semantics::Symbol *> &useDeviceAddrSyms, |
1850 | llvm::SmallVectorImpl<const semantics::Symbol *> &useDevicePtrSyms) { |
1851 | ClauseProcessor cp(converter, semaCtx, clauses); |
1852 | cp.processDevice(stmtCtx, clauseOps); |
1853 | cp.processIf(llvm::omp::Directive::OMPD_target_data, clauseOps); |
1854 | cp.processMap(loc, stmtCtx, clauseOps); |
1855 | cp.processUseDeviceAddr(stmtCtx, clauseOps, useDeviceAddrSyms); |
1856 | cp.processUseDevicePtr(stmtCtx, clauseOps, useDevicePtrSyms); |
1857 | |
1858 | // This function implements the deprecated functionality of use_device_ptr |
1859 | // that allows users to provide non-CPTR arguments to it with the caveat |
1860 | // that the compiler will treat them as use_device_addr. A lot of legacy |
1861 | // code may still depend on this functionality, so we should support it |
1862 | // in some manner. We do so currently by simply shifting non-cptr operands |
1863 | // from the use_device_ptr lists into the use_device_addr lists. |
1864 | // TODO: Perhaps create a user provideable compiler option that will |
1865 | // re-introduce a hard-error rather than a warning in these cases. |
1866 | promoteNonCPtrUseDevicePtrArgsToUseDeviceAddr( |
1867 | clauseOps.useDeviceAddrVars, useDeviceAddrSyms, |
1868 | clauseOps.useDevicePtrVars, useDevicePtrSyms); |
1869 | } |
1870 | |
1871 | static void genTargetEnterExitUpdateDataClauses( |
1872 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1873 | lower::SymMap &symTable, lower::StatementContext &stmtCtx, |
1874 | const List<Clause> &clauses, mlir::Location loc, |
1875 | llvm::omp::Directive directive, |
1876 | mlir::omp::TargetEnterExitUpdateDataOperands &clauseOps) { |
1877 | ClauseProcessor cp(converter, semaCtx, clauses); |
1878 | cp.processDepend(symTable, stmtCtx, clauseOps); |
1879 | cp.processDevice(stmtCtx, clauseOps); |
1880 | cp.processIf(directive, clauseOps); |
1881 | |
1882 | if (directive == llvm::omp::Directive::OMPD_target_update) |
1883 | cp.processMotionClauses(stmtCtx, clauseOps); |
1884 | else |
1885 | cp.processMap(loc, stmtCtx, clauseOps); |
1886 | |
1887 | cp.processNowait(clauseOps); |
1888 | } |
1889 | |
1890 | static void genTaskClauses( |
1891 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1892 | lower::SymMap &symTable, lower::StatementContext &stmtCtx, |
1893 | const List<Clause> &clauses, mlir::Location loc, |
1894 | mlir::omp::TaskOperands &clauseOps, |
1895 | llvm::SmallVectorImpl<const semantics::Symbol *> &inReductionSyms) { |
1896 | ClauseProcessor cp(converter, semaCtx, clauses); |
1897 | cp.processAllocate(clauseOps); |
1898 | cp.processDepend(symTable, stmtCtx, clauseOps); |
1899 | cp.processFinal(stmtCtx, clauseOps); |
1900 | cp.processIf(llvm::omp::Directive::OMPD_task, clauseOps); |
1901 | cp.processInReduction(loc, clauseOps, inReductionSyms); |
1902 | cp.processMergeable(clauseOps); |
1903 | cp.processPriority(stmtCtx, clauseOps); |
1904 | cp.processUntied(clauseOps); |
1905 | cp.processDetach(clauseOps); |
1906 | |
1907 | cp.processTODO<clause::Affinity>(loc, llvm::omp::Directive::OMPD_task); |
1908 | } |
1909 | |
1910 | static void genTaskgroupClauses( |
1911 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1912 | const List<Clause> &clauses, mlir::Location loc, |
1913 | mlir::omp::TaskgroupOperands &clauseOps, |
1914 | llvm::SmallVectorImpl<const semantics::Symbol *> &taskReductionSyms) { |
1915 | ClauseProcessor cp(converter, semaCtx, clauses); |
1916 | cp.processAllocate(clauseOps); |
1917 | cp.processTaskReduction(loc, clauseOps, taskReductionSyms); |
1918 | } |
1919 | |
1920 | static void genTaskloopClauses(lower::AbstractConverter &converter, |
1921 | semantics::SemanticsContext &semaCtx, |
1922 | lower::StatementContext &stmtCtx, |
1923 | const List<Clause> &clauses, mlir::Location loc, |
1924 | mlir::omp::TaskloopOperands &clauseOps) { |
1925 | |
1926 | ClauseProcessor cp(converter, semaCtx, clauses); |
1927 | cp.processGrainsize(stmtCtx, clauseOps); |
1928 | cp.processNumTasks(stmtCtx, clauseOps); |
1929 | |
1930 | cp.processTODO<clause::Allocate, clause::Collapse, clause::Default, |
1931 | clause::Final, clause::If, clause::InReduction, |
1932 | clause::Lastprivate, clause::Mergeable, clause::Nogroup, |
1933 | clause::Priority, clause::Reduction, clause::Shared, |
1934 | clause::Untied>(loc, llvm::omp::Directive::OMPD_taskloop); |
1935 | } |
1936 | |
1937 | static void genTaskwaitClauses(lower::AbstractConverter &converter, |
1938 | semantics::SemanticsContext &semaCtx, |
1939 | const List<Clause> &clauses, mlir::Location loc, |
1940 | mlir::omp::TaskwaitOperands &clauseOps) { |
1941 | ClauseProcessor cp(converter, semaCtx, clauses); |
1942 | cp.processTODO<clause::Depend, clause::Nowait>( |
1943 | loc, llvm::omp::Directive::OMPD_taskwait); |
1944 | } |
1945 | |
1946 | static void genWorkshareClauses(lower::AbstractConverter &converter, |
1947 | semantics::SemanticsContext &semaCtx, |
1948 | lower::StatementContext &stmtCtx, |
1949 | const List<Clause> &clauses, mlir::Location loc, |
1950 | mlir::omp::WorkshareOperands &clauseOps) { |
1951 | ClauseProcessor cp(converter, semaCtx, clauses); |
1952 | cp.processNowait(clauseOps); |
1953 | } |
1954 | |
1955 | static void genTeamsClauses( |
1956 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1957 | lower::StatementContext &stmtCtx, const List<Clause> &clauses, |
1958 | mlir::Location loc, mlir::omp::TeamsOperands &clauseOps, |
1959 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1960 | ClauseProcessor cp(converter, semaCtx, clauses); |
1961 | cp.processAllocate(clauseOps); |
1962 | cp.processIf(llvm::omp::Directive::OMPD_teams, clauseOps); |
1963 | |
1964 | if (hostEvalInfo.empty() || !hostEvalInfo.back().apply(clauseOps)) { |
1965 | cp.processNumTeams(stmtCtx, clauseOps); |
1966 | cp.processThreadLimit(stmtCtx, clauseOps); |
1967 | } |
1968 | |
1969 | cp.processReduction(loc, clauseOps, reductionSyms); |
1970 | // TODO Support delayed privatization. |
1971 | } |
1972 | |
1973 | static void genWsloopClauses( |
1974 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
1975 | lower::StatementContext &stmtCtx, const List<Clause> &clauses, |
1976 | mlir::Location loc, mlir::omp::WsloopOperands &clauseOps, |
1977 | llvm::SmallVectorImpl<const semantics::Symbol *> &reductionSyms) { |
1978 | ClauseProcessor cp(converter, semaCtx, clauses); |
1979 | cp.processNowait(clauseOps); |
1980 | cp.processLinear(clauseOps); |
1981 | cp.processOrder(clauseOps); |
1982 | cp.processOrdered(clauseOps); |
1983 | cp.processReduction(loc, clauseOps, reductionSyms); |
1984 | cp.processSchedule(stmtCtx, clauseOps); |
1985 | |
1986 | cp.processTODO<clause::Allocate>(loc, llvm::omp::Directive::OMPD_do); |
1987 | } |
1988 | |
1989 | //===----------------------------------------------------------------------===// |
1990 | // Code generation functions for leaf constructs |
1991 | //===----------------------------------------------------------------------===// |
1992 | |
1993 | static mlir::omp::BarrierOp |
1994 | genBarrierOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
1995 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
1996 | mlir::Location loc, const ConstructQueue &queue, |
1997 | ConstructQueue::const_iterator item) { |
1998 | return converter.getFirOpBuilder().create<mlir::omp::BarrierOp>(loc); |
1999 | } |
2000 | |
2001 | static mlir::omp::CancelOp genCancelOp(lower::AbstractConverter &converter, |
2002 | semantics::SemanticsContext &semaCtx, |
2003 | lower::pft::Evaluation &eval, |
2004 | mlir::Location loc, |
2005 | const ConstructQueue &queue, |
2006 | ConstructQueue::const_iterator item) { |
2007 | mlir::omp::CancelOperands clauseOps; |
2008 | genCancelClauses(converter, semaCtx, item->clauses, loc, clauseOps); |
2009 | |
2010 | return converter.getFirOpBuilder().create<mlir::omp::CancelOp>(loc, |
2011 | clauseOps); |
2012 | } |
2013 | |
2014 | static mlir::omp::CancellationPointOp genCancellationPointOp( |
2015 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
2016 | lower::pft::Evaluation &eval, mlir::Location loc, |
2017 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2018 | mlir::omp::CancellationPointOperands clauseOps; |
2019 | genCancellationPointClauses(converter, semaCtx, item->clauses, loc, |
2020 | clauseOps); |
2021 | |
2022 | return converter.getFirOpBuilder().create<mlir::omp::CancellationPointOp>( |
2023 | loc, clauseOps); |
2024 | } |
2025 | |
2026 | static mlir::omp::CriticalOp |
2027 | genCriticalOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2028 | semantics::SemanticsContext &semaCtx, |
2029 | lower::pft::Evaluation &eval, mlir::Location loc, |
2030 | const ConstructQueue &queue, ConstructQueue::const_iterator item, |
2031 | const std::optional<parser::Name> &name) { |
2032 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
2033 | mlir::FlatSymbolRefAttr nameAttr; |
2034 | |
2035 | if (name) { |
2036 | std::string nameStr = name->ToString(); |
2037 | mlir::ModuleOp mod = firOpBuilder.getModule(); |
2038 | auto global = mod.lookupSymbol<mlir::omp::CriticalDeclareOp>(nameStr); |
2039 | if (!global) { |
2040 | mlir::omp::CriticalDeclareOperands clauseOps; |
2041 | genCriticalDeclareClauses(converter, semaCtx, item->clauses, loc, |
2042 | clauseOps, nameStr); |
2043 | |
2044 | mlir::OpBuilder modBuilder(mod.getBodyRegion()); |
2045 | global = modBuilder.create<mlir::omp::CriticalDeclareOp>(loc, clauseOps); |
2046 | } |
2047 | nameAttr = mlir::FlatSymbolRefAttr::get(firOpBuilder.getContext(), |
2048 | global.getSymName()); |
2049 | } |
2050 | |
2051 | return genOpWithBody<mlir::omp::CriticalOp>( |
2052 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2053 | llvm::omp::Directive::OMPD_critical), |
2054 | queue, item, nameAttr); |
2055 | } |
2056 | |
2057 | static mlir::omp::FlushOp |
2058 | genFlushOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2059 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2060 | mlir::Location loc, const ObjectList &objects, |
2061 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2062 | llvm::SmallVector<mlir::Value> operandRange; |
2063 | genFlushClauses(converter, semaCtx, objects, item->clauses, loc, |
2064 | operandRange); |
2065 | |
2066 | return converter.getFirOpBuilder().create<mlir::omp::FlushOp>( |
2067 | converter.getCurrentLocation(), operandRange); |
2068 | } |
2069 | |
2070 | static mlir::omp::LoopNestOp genLoopNestOp( |
2071 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2072 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2073 | mlir::Location loc, const ConstructQueue &queue, |
2074 | ConstructQueue::const_iterator item, mlir::omp::LoopNestOperands &clauseOps, |
2075 | llvm::ArrayRef<const semantics::Symbol *> iv, |
2076 | llvm::ArrayRef< |
2077 | std::pair<mlir::omp::BlockArgOpenMPOpInterface, const EntryBlockArgs &>> |
2078 | wrapperArgs, |
2079 | llvm::omp::Directive directive, DataSharingProcessor &dsp) { |
2080 | auto ivCallback = [&](mlir::Operation *op) { |
2081 | genLoopVars(op, converter, loc, iv, wrapperArgs); |
2082 | return llvm::SmallVector<const semantics::Symbol *>(iv); |
2083 | }; |
2084 | |
2085 | auto *nestedEval = |
2086 | getCollapsedLoopEval(eval, collapseValue: getCollapseValue(clauses: item->clauses)); |
2087 | |
2088 | return genOpWithBody<mlir::omp::LoopNestOp>( |
2089 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, *nestedEval, |
2090 | directive) |
2091 | .setClauses(&item->clauses) |
2092 | .setDataSharingProcessor(&dsp) |
2093 | .setGenRegionEntryCb(ivCallback), |
2094 | queue, item, clauseOps); |
2095 | } |
2096 | |
2097 | static mlir::omp::LoopOp |
2098 | genLoopOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2099 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2100 | mlir::Location loc, const ConstructQueue &queue, |
2101 | ConstructQueue::const_iterator item) { |
2102 | mlir::omp::LoopOperands loopClauseOps; |
2103 | llvm::SmallVector<const semantics::Symbol *> loopReductionSyms; |
2104 | genLoopClauses(converter, semaCtx, item->clauses, loc, loopClauseOps, |
2105 | loopReductionSyms); |
2106 | |
2107 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2108 | /*shouldCollectPreDeterminedSymbols=*/true, |
2109 | /*useDelayedPrivatization=*/true, symTable); |
2110 | dsp.processStep1(clauseOps: &loopClauseOps); |
2111 | |
2112 | mlir::omp::LoopNestOperands loopNestClauseOps; |
2113 | llvm::SmallVector<const semantics::Symbol *> iv; |
2114 | genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc, |
2115 | loopNestClauseOps, iv); |
2116 | |
2117 | EntryBlockArgs loopArgs; |
2118 | loopArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
2119 | loopArgs.priv.vars = loopClauseOps.privateVars; |
2120 | loopArgs.reduction.syms = loopReductionSyms; |
2121 | loopArgs.reduction.vars = loopClauseOps.reductionVars; |
2122 | |
2123 | auto loopOp = |
2124 | genWrapperOp<mlir::omp::LoopOp>(converter, loc, loopClauseOps, loopArgs); |
2125 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item, |
2126 | loopNestClauseOps, iv, {{loopOp, loopArgs}}, |
2127 | llvm::omp::Directive::OMPD_loop, dsp); |
2128 | return loopOp; |
2129 | } |
2130 | |
2131 | static mlir::omp::MaskedOp |
2132 | genMaskedOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2133 | lower::StatementContext &stmtCtx, |
2134 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2135 | mlir::Location loc, const ConstructQueue &queue, |
2136 | ConstructQueue::const_iterator item) { |
2137 | mlir::omp::MaskedOperands clauseOps; |
2138 | genMaskedClauses(converter, semaCtx, stmtCtx, item->clauses, loc, clauseOps); |
2139 | |
2140 | return genOpWithBody<mlir::omp::MaskedOp>( |
2141 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2142 | llvm::omp::Directive::OMPD_masked), |
2143 | queue, item, clauseOps); |
2144 | } |
2145 | |
2146 | static mlir::omp::MasterOp |
2147 | genMasterOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2148 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2149 | mlir::Location loc, const ConstructQueue &queue, |
2150 | ConstructQueue::const_iterator item) { |
2151 | return genOpWithBody<mlir::omp::MasterOp>( |
2152 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2153 | llvm::omp::Directive::OMPD_master), |
2154 | queue, item); |
2155 | } |
2156 | |
2157 | static mlir::omp::OrderedOp |
2158 | genOrderedOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2159 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2160 | mlir::Location loc, const ConstructQueue &queue, |
2161 | ConstructQueue::const_iterator item) { |
2162 | TODO(loc, "OMPD_ordered" ); |
2163 | return nullptr; |
2164 | } |
2165 | |
2166 | static mlir::omp::OrderedRegionOp |
2167 | genOrderedRegionOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2168 | semantics::SemanticsContext &semaCtx, |
2169 | lower::pft::Evaluation &eval, mlir::Location loc, |
2170 | const ConstructQueue &queue, |
2171 | ConstructQueue::const_iterator item) { |
2172 | mlir::omp::OrderedRegionOperands clauseOps; |
2173 | genOrderedRegionClauses(converter, semaCtx, item->clauses, loc, clauseOps); |
2174 | |
2175 | return genOpWithBody<mlir::omp::OrderedRegionOp>( |
2176 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2177 | llvm::omp::Directive::OMPD_ordered), |
2178 | queue, item, clauseOps); |
2179 | } |
2180 | |
2181 | static mlir::omp::ParallelOp |
2182 | genParallelOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2183 | semantics::SemanticsContext &semaCtx, |
2184 | lower::pft::Evaluation &eval, mlir::Location loc, |
2185 | const ConstructQueue &queue, ConstructQueue::const_iterator item, |
2186 | mlir::omp::ParallelOperands &clauseOps, |
2187 | const EntryBlockArgs &args, DataSharingProcessor *dsp, |
2188 | bool isComposite = false) { |
2189 | assert((!enableDelayedPrivatization || dsp) && |
2190 | "expected valid DataSharingProcessor" ); |
2191 | |
2192 | OpWithBodyGenInfo genInfo = |
2193 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2194 | llvm::omp::Directive::OMPD_parallel) |
2195 | .setClauses(&item->clauses) |
2196 | .setEntryBlockArgs(&args) |
2197 | .setGenSkeletonOnly(isComposite) |
2198 | .setDataSharingProcessor(dsp); |
2199 | |
2200 | auto parallelOp = |
2201 | genOpWithBody<mlir::omp::ParallelOp>(genInfo, queue, item, clauseOps); |
2202 | parallelOp.setComposite(isComposite); |
2203 | return parallelOp; |
2204 | } |
2205 | |
2206 | static mlir::omp::ScanOp |
2207 | genScanOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2208 | semantics::SemanticsContext &semaCtx, mlir::Location loc, |
2209 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2210 | mlir::omp::ScanOperands clauseOps; |
2211 | genScanClauses(converter, semaCtx, item->clauses, loc, clauseOps); |
2212 | return converter.getFirOpBuilder().create<mlir::omp::ScanOp>( |
2213 | converter.getCurrentLocation(), clauseOps); |
2214 | } |
2215 | |
2216 | /// This breaks the normal prototype of the gen*Op functions: adding the |
2217 | /// sectionBlocks argument so that the enclosed section constructs can be |
2218 | /// lowered here with correct reduction symbol remapping. |
2219 | static mlir::omp::SectionsOp |
2220 | genSectionsOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2221 | semantics::SemanticsContext &semaCtx, |
2222 | lower::pft::Evaluation &eval, mlir::Location loc, |
2223 | const ConstructQueue &queue, ConstructQueue::const_iterator item, |
2224 | const parser::OmpSectionBlocks §ionBlocks) { |
2225 | mlir::omp::SectionsOperands clauseOps; |
2226 | llvm::SmallVector<const semantics::Symbol *> reductionSyms; |
2227 | genSectionsClauses(converter, semaCtx, item->clauses, loc, clauseOps, |
2228 | reductionSyms); |
2229 | |
2230 | auto &builder = converter.getFirOpBuilder(); |
2231 | |
2232 | // Insert privatizations before SECTIONS |
2233 | lower::SymMapScope scope(symTable); |
2234 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2235 | lower::omp::isLastItemInQueue(item, queue), |
2236 | /*useDelayedPrivatization=*/false, symTable); |
2237 | dsp.processStep1(); |
2238 | |
2239 | List<Clause> nonDsaClauses; |
2240 | List<const clause::Lastprivate *> lastprivates; |
2241 | |
2242 | for (const Clause &clause : item->clauses) { |
2243 | if (clause.id == llvm::omp::Clause::OMPC_lastprivate) { |
2244 | auto &lastp = std::get<clause::Lastprivate>(clause.u); |
2245 | lastprivateModifierNotSupported(lastp, converter.getCurrentLocation()); |
2246 | lastprivates.push_back(&lastp); |
2247 | } else { |
2248 | switch (clause.id) { |
2249 | case llvm::omp::Clause::OMPC_firstprivate: |
2250 | case llvm::omp::Clause::OMPC_private: |
2251 | case llvm::omp::Clause::OMPC_shared: |
2252 | break; |
2253 | default: |
2254 | nonDsaClauses.push_back(Elt: clause); |
2255 | } |
2256 | } |
2257 | } |
2258 | |
2259 | // SECTIONS construct. |
2260 | auto sectionsOp = builder.create<mlir::omp::SectionsOp>(loc, clauseOps); |
2261 | |
2262 | // Create entry block with reduction variables as arguments. |
2263 | EntryBlockArgs args; |
2264 | // TODO: Add private syms and vars. |
2265 | args.reduction.syms = reductionSyms; |
2266 | args.reduction.vars = clauseOps.reductionVars; |
2267 | |
2268 | genEntryBlock(builder, args, sectionsOp.getRegion()); |
2269 | mlir::Operation *terminator = |
2270 | lower::genOpenMPTerminator(builder, sectionsOp, loc); |
2271 | |
2272 | // Generate nested SECTION constructs. |
2273 | // This is done here rather than in genOMP([...], OpenMPSectionConstruct ) |
2274 | // because we need to run genReductionVars on each omp.section so that the |
2275 | // reduction variable gets mapped to the private version |
2276 | for (auto [construct, nestedEval] : |
2277 | llvm::zip(sectionBlocks.v, eval.getNestedEvaluations())) { |
2278 | const auto *sectionConstruct = |
2279 | std::get_if<parser::OpenMPSectionConstruct>(&construct.u); |
2280 | if (!sectionConstruct) { |
2281 | assert(false && |
2282 | "unexpected construct nested inside of SECTIONS construct" ); |
2283 | continue; |
2284 | } |
2285 | |
2286 | ConstructQueue sectionQueue{buildConstructQueue( |
2287 | converter.getFirOpBuilder().getModule(), semaCtx, nestedEval, |
2288 | sectionConstruct->source, llvm::omp::Directive::OMPD_section, {})}; |
2289 | |
2290 | builder.setInsertionPoint(terminator); |
2291 | genOpWithBody<mlir::omp::SectionOp>( |
2292 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, nestedEval, |
2293 | llvm::omp::Directive::OMPD_section) |
2294 | .setClauses(§ionQueue.begin()->clauses) |
2295 | .setDataSharingProcessor(&dsp) |
2296 | .setEntryBlockArgs(&args), |
2297 | sectionQueue, sectionQueue.begin()); |
2298 | } |
2299 | |
2300 | if (!lastprivates.empty()) { |
2301 | mlir::Region §ionsBody = sectionsOp.getRegion(); |
2302 | assert(sectionsBody.hasOneBlock()); |
2303 | mlir::Block &body = sectionsBody.front(); |
2304 | |
2305 | auto lastSectionOp = llvm::find_if( |
2306 | llvm::reverse(body.getOperations()), [](const mlir::Operation &op) { |
2307 | return llvm::isa<mlir::omp::SectionOp>(op); |
2308 | }); |
2309 | assert(lastSectionOp != body.rend()); |
2310 | |
2311 | for (const clause::Lastprivate *lastp : lastprivates) { |
2312 | builder.setInsertionPoint( |
2313 | lastSectionOp->getRegion(0).back().getTerminator()); |
2314 | mlir::OpBuilder::InsertPoint insp = builder.saveInsertionPoint(); |
2315 | const auto &objList = std::get<ObjectList>(lastp->t); |
2316 | for (const Object &object : objList) { |
2317 | semantics::Symbol *sym = object.sym(); |
2318 | if (const auto *common = |
2319 | sym->detailsIf<semantics::CommonBlockDetails>()) { |
2320 | for (const auto &obj : common->objects()) |
2321 | converter.copyHostAssociateVar(*obj, &insp, /*hostIsSource=*/false); |
2322 | } else { |
2323 | converter.copyHostAssociateVar(*sym, &insp, /*hostIsSource=*/false); |
2324 | } |
2325 | } |
2326 | } |
2327 | } |
2328 | |
2329 | // Perform DataSharingProcessor's step2 out of SECTIONS |
2330 | builder.setInsertionPointAfter(sectionsOp.getOperation()); |
2331 | dsp.processStep2(sectionsOp, false); |
2332 | // Emit implicit barrier to synchronize threads and avoid data |
2333 | // races on post-update of lastprivate variables when `nowait` |
2334 | // clause is present. |
2335 | if (clauseOps.nowait && !lastprivates.empty()) |
2336 | builder.create<mlir::omp::BarrierOp>(loc); |
2337 | |
2338 | return sectionsOp; |
2339 | } |
2340 | |
2341 | static mlir::Operation * |
2342 | genScopeOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2343 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2344 | mlir::Location loc, const ConstructQueue &queue, |
2345 | ConstructQueue::const_iterator item) { |
2346 | TODO(loc, "Scope construct" ); |
2347 | return nullptr; |
2348 | } |
2349 | |
2350 | static mlir::omp::SingleOp |
2351 | genSingleOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2352 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2353 | mlir::Location loc, const ConstructQueue &queue, |
2354 | ConstructQueue::const_iterator item) { |
2355 | mlir::omp::SingleOperands clauseOps; |
2356 | genSingleClauses(converter, semaCtx, item->clauses, loc, clauseOps); |
2357 | |
2358 | return genOpWithBody<mlir::omp::SingleOp>( |
2359 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2360 | llvm::omp::Directive::OMPD_single) |
2361 | .setClauses(&item->clauses), |
2362 | queue, item, clauseOps); |
2363 | } |
2364 | |
2365 | static mlir::omp::TargetOp |
2366 | genTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2367 | lower::StatementContext &stmtCtx, |
2368 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2369 | mlir::Location loc, const ConstructQueue &queue, |
2370 | ConstructQueue::const_iterator item) { |
2371 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
2372 | bool isTargetDevice = |
2373 | llvm::cast<mlir::omp::OffloadModuleInterface>(*converter.getModuleOp()) |
2374 | .getIsTargetDevice(); |
2375 | |
2376 | // Introduce a new host_eval information structure for this target region. |
2377 | if (!isTargetDevice) |
2378 | hostEvalInfo.emplace_back(); |
2379 | |
2380 | mlir::omp::TargetOperands clauseOps; |
2381 | DefaultMapsTy defaultMaps; |
2382 | llvm::SmallVector<const semantics::Symbol *> mapSyms, isDevicePtrSyms, |
2383 | hasDeviceAddrSyms; |
2384 | genTargetClauses(converter, semaCtx, symTable, stmtCtx, eval, item->clauses, |
2385 | loc, clauseOps, defaultMaps, hasDeviceAddrSyms, |
2386 | isDevicePtrSyms, mapSyms); |
2387 | |
2388 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2389 | /*shouldCollectPreDeterminedSymbols=*/ |
2390 | lower::omp::isLastItemInQueue(item, queue), |
2391 | /*useDelayedPrivatization=*/true, symTable); |
2392 | dsp.processStep1(clauseOps: &clauseOps); |
2393 | |
2394 | // 5.8.1 Implicit Data-Mapping Attribute Rules |
2395 | // The following code follows the implicit data-mapping rules to map all the |
2396 | // symbols used inside the region that do not have explicit data-environment |
2397 | // attribute clauses (neither data-sharing; e.g. `private`, nor `map` |
2398 | // clauses). |
2399 | auto captureImplicitMap = [&](const semantics::Symbol &sym) { |
2400 | if (dsp.getAllSymbolsToPrivatize().contains(&sym)) |
2401 | return; |
2402 | |
2403 | // These symbols are mapped individually in processHasDeviceAddr. |
2404 | if (llvm::is_contained(Range&: hasDeviceAddrSyms, Element: &sym)) |
2405 | return; |
2406 | |
2407 | // Structure component symbols don't have bindings, and can only be |
2408 | // explicitly mapped individually. If a member is captured implicitly |
2409 | // we map the entirety of the derived type when we find its symbol. |
2410 | if (sym.owner().IsDerivedType()) |
2411 | return; |
2412 | |
2413 | // if the symbol is part of an already mapped common block, do not make a |
2414 | // map for it. |
2415 | if (const Fortran::semantics::Symbol *common = |
2416 | Fortran::semantics::FindCommonBlockContaining(sym.GetUltimate())) |
2417 | if (llvm::is_contained(Range&: mapSyms, Element: common)) |
2418 | return; |
2419 | |
2420 | // If we come across a symbol without a symbol address, we |
2421 | // return as we cannot process it, this is intended as a |
2422 | // catch all early exit for symbols that do not have a |
2423 | // corresponding extended value. Such as subroutines, |
2424 | // interfaces and named blocks. |
2425 | if (!converter.getSymbolAddress(sym)) |
2426 | return; |
2427 | |
2428 | if (!llvm::is_contained(Range&: mapSyms, Element: &sym)) { |
2429 | if (const auto *details = |
2430 | sym.template detailsIf<semantics::HostAssocDetails>()) |
2431 | converter.copySymbolBinding(details->symbol(), sym); |
2432 | std::stringstream name; |
2433 | fir::ExtendedValue dataExv = converter.getSymbolExtendedValue(sym); |
2434 | name << sym.name().ToString(); |
2435 | |
2436 | mlir::FlatSymbolRefAttr mapperId; |
2437 | if (sym.GetType()->category() == semantics::DeclTypeSpec::TypeDerived) { |
2438 | auto &typeSpec = sym.GetType()->derivedTypeSpec(); |
2439 | std::string mapperIdName = |
2440 | typeSpec.name().ToString() + llvm::omp::OmpDefaultMapperName; |
2441 | if (auto *sym = converter.getCurrentScope().FindSymbol(mapperIdName)) |
2442 | mapperIdName = converter.mangleName(mapperIdName, sym->owner()); |
2443 | if (converter.getModuleOp().lookupSymbol(mapperIdName)) |
2444 | mapperId = mlir::FlatSymbolRefAttr::get(&converter.getMLIRContext(), |
2445 | mapperIdName); |
2446 | } |
2447 | |
2448 | fir::factory::AddrAndBoundsInfo info = |
2449 | Fortran::lower::getDataOperandBaseAddr( |
2450 | converter, firOpBuilder, sym.GetUltimate(), |
2451 | converter.getCurrentLocation()); |
2452 | llvm::SmallVector<mlir::Value> bounds = |
2453 | fir::factory::genImplicitBoundsOps<mlir::omp::MapBoundsOp, |
2454 | mlir::omp::MapBoundsType>( |
2455 | firOpBuilder, info, dataExv, |
2456 | semantics::IsAssumedSizeArray(sym.GetUltimate()), |
2457 | converter.getCurrentLocation()); |
2458 | mlir::Value baseOp = info.rawInput; |
2459 | mlir::Type eleType = baseOp.getType(); |
2460 | if (auto refType = mlir::dyn_cast<fir::ReferenceType>(baseOp.getType())) |
2461 | eleType = refType.getElementType(); |
2462 | |
2463 | std::pair<llvm::omp::OpenMPOffloadMappingFlags, |
2464 | mlir::omp::VariableCaptureKind> |
2465 | mapFlagAndKind = getImplicitMapTypeAndKind( |
2466 | firOpBuilder, converter, defaultMaps, eleType, loc, sym); |
2467 | |
2468 | mlir::Value mapOp = createMapInfoOp( |
2469 | firOpBuilder, converter.getCurrentLocation(), baseOp, |
2470 | /*varPtrPtr=*/mlir::Value{}, name.str(), bounds, /*members=*/{}, |
2471 | /*membersIndex=*/mlir::ArrayAttr{}, |
2472 | static_cast< |
2473 | std::underlying_type_t<llvm::omp::OpenMPOffloadMappingFlags>>( |
2474 | std::get<0>(mapFlagAndKind)), |
2475 | std::get<1>(mapFlagAndKind), baseOp.getType(), |
2476 | /*partialMap=*/false, mapperId); |
2477 | |
2478 | clauseOps.mapVars.push_back(mapOp); |
2479 | mapSyms.push_back(Elt: &sym); |
2480 | } |
2481 | }; |
2482 | lower::pft::visitAllSymbols(eval, captureImplicitMap); |
2483 | |
2484 | auto targetOp = firOpBuilder.create<mlir::omp::TargetOp>(loc, clauseOps); |
2485 | |
2486 | llvm::SmallVector<mlir::Value> hasDeviceAddrBaseValues, mapBaseValues; |
2487 | extractMappedBaseValues(clauseOps.hasDeviceAddrVars, hasDeviceAddrBaseValues); |
2488 | extractMappedBaseValues(clauseOps.mapVars, mapBaseValues); |
2489 | |
2490 | EntryBlockArgs args; |
2491 | args.hasDeviceAddr.syms = hasDeviceAddrSyms; |
2492 | args.hasDeviceAddr.vars = hasDeviceAddrBaseValues; |
2493 | args.hostEvalVars = clauseOps.hostEvalVars; |
2494 | // TODO: Add in_reduction syms and vars. |
2495 | args.map.syms = mapSyms; |
2496 | args.map.vars = mapBaseValues; |
2497 | args.priv.syms = dsp.getDelayedPrivSymbols(); |
2498 | args.priv.vars = clauseOps.privateVars; |
2499 | |
2500 | genBodyOfTargetOp(converter, symTable, semaCtx, eval, targetOp, args, loc, |
2501 | queue, item, dsp); |
2502 | |
2503 | // Remove the host_eval information structure created for this target region. |
2504 | if (!isTargetDevice) |
2505 | hostEvalInfo.pop_back(); |
2506 | return targetOp; |
2507 | } |
2508 | |
2509 | static mlir::omp::TargetDataOp genTargetDataOp( |
2510 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2511 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2512 | lower::pft::Evaluation &eval, mlir::Location loc, |
2513 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2514 | mlir::omp::TargetDataOperands clauseOps; |
2515 | llvm::SmallVector<const semantics::Symbol *> useDeviceAddrSyms, |
2516 | useDevicePtrSyms; |
2517 | genTargetDataClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
2518 | clauseOps, useDeviceAddrSyms, useDevicePtrSyms); |
2519 | |
2520 | auto targetDataOp = |
2521 | converter.getFirOpBuilder().create<mlir::omp::TargetDataOp>(loc, |
2522 | clauseOps); |
2523 | |
2524 | llvm::SmallVector<mlir::Value> useDeviceAddrBaseValues, |
2525 | useDevicePtrBaseValues; |
2526 | extractMappedBaseValues(clauseOps.useDeviceAddrVars, useDeviceAddrBaseValues); |
2527 | extractMappedBaseValues(clauseOps.useDevicePtrVars, useDevicePtrBaseValues); |
2528 | |
2529 | EntryBlockArgs args; |
2530 | args.useDeviceAddr.syms = useDeviceAddrSyms; |
2531 | args.useDeviceAddr.vars = useDeviceAddrBaseValues; |
2532 | args.useDevicePtr.syms = useDevicePtrSyms; |
2533 | args.useDevicePtr.vars = useDevicePtrBaseValues; |
2534 | |
2535 | genBodyOfTargetDataOp(converter, symTable, semaCtx, eval, targetDataOp, args, |
2536 | loc, queue, item); |
2537 | return targetDataOp; |
2538 | } |
2539 | |
2540 | template <typename OpTy> |
2541 | static OpTy genTargetEnterExitUpdateDataOp( |
2542 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2543 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2544 | mlir::Location loc, const ConstructQueue &queue, |
2545 | ConstructQueue::const_iterator item) { |
2546 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
2547 | |
2548 | // GCC 9.3.0 emits a (probably) bogus warning about an unused variable. |
2549 | [[maybe_unused]] llvm::omp::Directive directive; |
2550 | if constexpr (std::is_same_v<OpTy, mlir::omp::TargetEnterDataOp>) { |
2551 | directive = llvm::omp::Directive::OMPD_target_enter_data; |
2552 | } else if constexpr (std::is_same_v<OpTy, mlir::omp::TargetExitDataOp>) { |
2553 | directive = llvm::omp::Directive::OMPD_target_exit_data; |
2554 | } else if constexpr (std::is_same_v<OpTy, mlir::omp::TargetUpdateOp>) { |
2555 | directive = llvm::omp::Directive::OMPD_target_update; |
2556 | } else { |
2557 | llvm_unreachable("Unexpected TARGET DATA construct" ); |
2558 | } |
2559 | |
2560 | mlir::omp::TargetEnterExitUpdateDataOperands clauseOps; |
2561 | genTargetEnterExitUpdateDataClauses(converter, semaCtx, symTable, stmtCtx, |
2562 | item->clauses, loc, directive, clauseOps); |
2563 | |
2564 | return firOpBuilder.create<OpTy>(loc, clauseOps); |
2565 | } |
2566 | |
2567 | static mlir::omp::TaskOp |
2568 | genTaskOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2569 | lower::StatementContext &stmtCtx, |
2570 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2571 | mlir::Location loc, const ConstructQueue &queue, |
2572 | ConstructQueue::const_iterator item) { |
2573 | mlir::omp::TaskOperands clauseOps; |
2574 | llvm::SmallVector<const semantics::Symbol *> inReductionSyms; |
2575 | genTaskClauses(converter, semaCtx, symTable, stmtCtx, item->clauses, loc, |
2576 | clauseOps, inReductionSyms); |
2577 | |
2578 | if (!enableDelayedPrivatization) |
2579 | return genOpWithBody<mlir::omp::TaskOp>( |
2580 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2581 | llvm::omp::Directive::OMPD_task) |
2582 | .setClauses(&item->clauses), |
2583 | queue, item, clauseOps); |
2584 | |
2585 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2586 | lower::omp::isLastItemInQueue(item, queue), |
2587 | /*useDelayedPrivatization=*/true, symTable); |
2588 | dsp.processStep1(clauseOps: &clauseOps); |
2589 | |
2590 | EntryBlockArgs taskArgs; |
2591 | taskArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
2592 | taskArgs.priv.vars = clauseOps.privateVars; |
2593 | taskArgs.inReduction.syms = inReductionSyms; |
2594 | taskArgs.inReduction.vars = clauseOps.inReductionVars; |
2595 | |
2596 | return genOpWithBody<mlir::omp::TaskOp>( |
2597 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2598 | llvm::omp::Directive::OMPD_task) |
2599 | .setClauses(&item->clauses) |
2600 | .setDataSharingProcessor(&dsp) |
2601 | .setEntryBlockArgs(&taskArgs), |
2602 | queue, item, clauseOps); |
2603 | } |
2604 | |
2605 | static mlir::omp::TaskgroupOp |
2606 | genTaskgroupOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2607 | semantics::SemanticsContext &semaCtx, |
2608 | lower::pft::Evaluation &eval, mlir::Location loc, |
2609 | const ConstructQueue &queue, |
2610 | ConstructQueue::const_iterator item) { |
2611 | mlir::omp::TaskgroupOperands clauseOps; |
2612 | llvm::SmallVector<const semantics::Symbol *> taskReductionSyms; |
2613 | genTaskgroupClauses(converter, semaCtx, item->clauses, loc, clauseOps, |
2614 | taskReductionSyms); |
2615 | |
2616 | EntryBlockArgs taskgroupArgs; |
2617 | taskgroupArgs.taskReduction.syms = taskReductionSyms; |
2618 | taskgroupArgs.taskReduction.vars = clauseOps.taskReductionVars; |
2619 | |
2620 | return genOpWithBody<mlir::omp::TaskgroupOp>( |
2621 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2622 | llvm::omp::Directive::OMPD_taskgroup) |
2623 | .setClauses(&item->clauses) |
2624 | .setEntryBlockArgs(&taskgroupArgs), |
2625 | queue, item, clauseOps); |
2626 | } |
2627 | |
2628 | static mlir::omp::TaskwaitOp |
2629 | genTaskwaitOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2630 | semantics::SemanticsContext &semaCtx, |
2631 | lower::pft::Evaluation &eval, mlir::Location loc, |
2632 | const ConstructQueue &queue, |
2633 | ConstructQueue::const_iterator item) { |
2634 | mlir::omp::TaskwaitOperands clauseOps; |
2635 | genTaskwaitClauses(converter, semaCtx, item->clauses, loc, clauseOps); |
2636 | return converter.getFirOpBuilder().create<mlir::omp::TaskwaitOp>(loc, |
2637 | clauseOps); |
2638 | } |
2639 | |
2640 | static mlir::omp::TaskyieldOp |
2641 | genTaskyieldOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2642 | semantics::SemanticsContext &semaCtx, |
2643 | lower::pft::Evaluation &eval, mlir::Location loc, |
2644 | const ConstructQueue &queue, |
2645 | ConstructQueue::const_iterator item) { |
2646 | return converter.getFirOpBuilder().create<mlir::omp::TaskyieldOp>(loc); |
2647 | } |
2648 | |
2649 | static mlir::omp::WorkshareOp genWorkshareOp( |
2650 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2651 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2652 | lower::pft::Evaluation &eval, mlir::Location loc, |
2653 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2654 | mlir::omp::WorkshareOperands clauseOps; |
2655 | genWorkshareClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
2656 | clauseOps); |
2657 | |
2658 | return genOpWithBody<mlir::omp::WorkshareOp>( |
2659 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2660 | llvm::omp::Directive::OMPD_workshare) |
2661 | .setClauses(&item->clauses), |
2662 | queue, item, clauseOps); |
2663 | } |
2664 | |
2665 | static mlir::omp::TeamsOp |
2666 | genTeamsOp(lower::AbstractConverter &converter, lower::SymMap &symTable, |
2667 | lower::StatementContext &stmtCtx, |
2668 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
2669 | mlir::Location loc, const ConstructQueue &queue, |
2670 | ConstructQueue::const_iterator item) { |
2671 | mlir::omp::TeamsOperands clauseOps; |
2672 | llvm::SmallVector<const semantics::Symbol *> reductionSyms; |
2673 | genTeamsClauses(converter, semaCtx, stmtCtx, item->clauses, loc, clauseOps, |
2674 | reductionSyms); |
2675 | |
2676 | EntryBlockArgs args; |
2677 | // TODO: Add private syms and vars. |
2678 | args.reduction.syms = reductionSyms; |
2679 | args.reduction.vars = clauseOps.reductionVars; |
2680 | |
2681 | return genOpWithBody<mlir::omp::TeamsOp>( |
2682 | OpWithBodyGenInfo(converter, symTable, semaCtx, loc, eval, |
2683 | llvm::omp::Directive::OMPD_teams) |
2684 | .setClauses(&item->clauses) |
2685 | .setEntryBlockArgs(&args), |
2686 | queue, item, clauseOps); |
2687 | } |
2688 | |
2689 | //===----------------------------------------------------------------------===// |
2690 | // Code generation for atomic operations |
2691 | //===----------------------------------------------------------------------===// |
2692 | static fir::FirOpBuilder::InsertPoint |
2693 | getInsertionPointBefore(mlir::Operation *op) { |
2694 | return fir::FirOpBuilder::InsertPoint(op->getBlock(), |
2695 | mlir::Block::iterator(op)); |
2696 | } |
2697 | |
2698 | static fir::FirOpBuilder::InsertPoint |
2699 | getInsertionPointAfter(mlir::Operation *op) { |
2700 | return fir::FirOpBuilder::InsertPoint(op->getBlock(), |
2701 | ++mlir::Block::iterator(op)); |
2702 | } |
2703 | |
2704 | static mlir::IntegerAttr getAtomicHint(lower::AbstractConverter &converter, |
2705 | const List<Clause> &clauses) { |
2706 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
2707 | for (const Clause &clause : clauses) { |
2708 | if (clause.id != llvm::omp::Clause::OMPC_hint) |
2709 | continue; |
2710 | auto &hint = std::get<clause::Hint>(clause.u); |
2711 | auto maybeVal = evaluate::ToInt64(hint.v); |
2712 | CHECK(maybeVal); |
2713 | return builder.getI64IntegerAttr(*maybeVal); |
2714 | } |
2715 | return nullptr; |
2716 | } |
2717 | |
2718 | static mlir::omp::ClauseMemoryOrderKindAttr |
2719 | getAtomicMemoryOrder(lower::AbstractConverter &converter, |
2720 | semantics::SemanticsContext &semaCtx, |
2721 | const List<Clause> &clauses) { |
2722 | std::optional<mlir::omp::ClauseMemoryOrderKind> kind; |
2723 | unsigned version = semaCtx.langOptions().OpenMPVersion; |
2724 | |
2725 | for (const Clause &clause : clauses) { |
2726 | switch (clause.id) { |
2727 | case llvm::omp::Clause::OMPC_acq_rel: |
2728 | kind = mlir::omp::ClauseMemoryOrderKind::Acq_rel; |
2729 | break; |
2730 | case llvm::omp::Clause::OMPC_acquire: |
2731 | kind = mlir::omp::ClauseMemoryOrderKind::Acquire; |
2732 | break; |
2733 | case llvm::omp::Clause::OMPC_relaxed: |
2734 | kind = mlir::omp::ClauseMemoryOrderKind::Relaxed; |
2735 | break; |
2736 | case llvm::omp::Clause::OMPC_release: |
2737 | kind = mlir::omp::ClauseMemoryOrderKind::Release; |
2738 | break; |
2739 | case llvm::omp::Clause::OMPC_seq_cst: |
2740 | kind = mlir::omp::ClauseMemoryOrderKind::Seq_cst; |
2741 | break; |
2742 | default: |
2743 | break; |
2744 | } |
2745 | } |
2746 | |
2747 | // Starting with 5.1, if no memory-order clause is present, the effect |
2748 | // is as if "relaxed" was present. |
2749 | if (!kind) { |
2750 | if (version <= 50) |
2751 | return nullptr; |
2752 | kind = mlir::omp::ClauseMemoryOrderKind::Relaxed; |
2753 | } |
2754 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
2755 | return mlir::omp::ClauseMemoryOrderKindAttr::get(builder.getContext(), *kind); |
2756 | } |
2757 | |
2758 | static mlir::Operation * // |
2759 | genAtomicRead(lower::AbstractConverter &converter, mlir::Location loc, |
2760 | lower::StatementContext &stmtCtx, mlir::Value atomAddr, |
2761 | const semantics::SomeExpr &atom, |
2762 | const evaluate::Assignment &assign, mlir::IntegerAttr hint, |
2763 | mlir::omp::ClauseMemoryOrderKindAttr memOrder, |
2764 | fir::FirOpBuilder::InsertPoint preAt, |
2765 | fir::FirOpBuilder::InsertPoint atomicAt, |
2766 | fir::FirOpBuilder::InsertPoint postAt) { |
2767 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
2768 | builder.restoreInsertionPoint(preAt); |
2769 | |
2770 | mlir::Value storeAddr = |
2771 | fir::getBase(converter.genExprAddr(assign.lhs, stmtCtx, &loc)); |
2772 | mlir::Type atomType = fir::unwrapRefType(atomAddr.getType()); |
2773 | mlir::Type storeType = fir::unwrapRefType(storeAddr.getType()); |
2774 | |
2775 | mlir::Value toAddr = [&]() { |
2776 | if (atomType == storeType) |
2777 | return storeAddr; |
2778 | return builder.createTemporary(loc, atomType, ".tmp.atomval" ); |
2779 | }(); |
2780 | |
2781 | builder.restoreInsertionPoint(atomicAt); |
2782 | mlir::Operation *op = builder.create<mlir::omp::AtomicReadOp>( |
2783 | loc, atomAddr, toAddr, mlir::TypeAttr::get(atomType), hint, memOrder); |
2784 | |
2785 | if (atomType != storeType) { |
2786 | lower::ExprToValueMap overrides; |
2787 | // The READ operation could be a part of UPDATE CAPTURE, so make sure |
2788 | // we don't emit extra code into the body of the atomic op. |
2789 | builder.restoreInsertionPoint(postAt); |
2790 | mlir::Value load = builder.create<fir::LoadOp>(loc, toAddr); |
2791 | overrides.try_emplace(&atom, load); |
2792 | |
2793 | converter.overrideExprValues(&overrides); |
2794 | mlir::Value value = |
2795 | fir::getBase(converter.genExprValue(assign.rhs, stmtCtx, &loc)); |
2796 | converter.resetExprOverrides(); |
2797 | |
2798 | builder.create<fir::StoreOp>(loc, value, storeAddr); |
2799 | } |
2800 | return op; |
2801 | } |
2802 | |
2803 | static mlir::Operation * // |
2804 | genAtomicWrite(lower::AbstractConverter &converter, mlir::Location loc, |
2805 | lower::StatementContext &stmtCtx, mlir::Value atomAddr, |
2806 | const semantics::SomeExpr &atom, |
2807 | const evaluate::Assignment &assign, mlir::IntegerAttr hint, |
2808 | mlir::omp::ClauseMemoryOrderKindAttr memOrder, |
2809 | fir::FirOpBuilder::InsertPoint preAt, |
2810 | fir::FirOpBuilder::InsertPoint atomicAt, |
2811 | fir::FirOpBuilder::InsertPoint postAt) { |
2812 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
2813 | builder.restoreInsertionPoint(preAt); |
2814 | |
2815 | mlir::Value value = |
2816 | fir::getBase(converter.genExprValue(assign.rhs, stmtCtx, &loc)); |
2817 | mlir::Type atomType = fir::unwrapRefType(atomAddr.getType()); |
2818 | mlir::Value converted = builder.createConvert(loc, atomType, value); |
2819 | |
2820 | builder.restoreInsertionPoint(atomicAt); |
2821 | mlir::Operation *op = builder.create<mlir::omp::AtomicWriteOp>( |
2822 | loc, atomAddr, converted, hint, memOrder); |
2823 | return op; |
2824 | } |
2825 | |
2826 | static mlir::Operation * |
2827 | genAtomicUpdate(lower::AbstractConverter &converter, mlir::Location loc, |
2828 | lower::StatementContext &stmtCtx, mlir::Value atomAddr, |
2829 | const semantics::SomeExpr &atom, |
2830 | const evaluate::Assignment &assign, mlir::IntegerAttr hint, |
2831 | mlir::omp::ClauseMemoryOrderKindAttr memOrder, |
2832 | fir::FirOpBuilder::InsertPoint preAt, |
2833 | fir::FirOpBuilder::InsertPoint atomicAt, |
2834 | fir::FirOpBuilder::InsertPoint postAt) { |
2835 | lower::ExprToValueMap overrides; |
2836 | lower::StatementContext naCtx; |
2837 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
2838 | builder.restoreInsertionPoint(preAt); |
2839 | |
2840 | mlir::Type atomType = fir::unwrapRefType(atomAddr.getType()); |
2841 | |
2842 | // This must exist by now. |
2843 | SomeExpr input = *semantics::GetConvertInput(assign.rhs); |
2844 | std::vector<SomeExpr> args{semantics::GetTopLevelOperation(input).second}; |
2845 | assert(!args.empty() && "Update operation without arguments" ); |
2846 | for (auto &arg : args) { |
2847 | if (!semantics::IsSameOrConvertOf(arg, atom)) { |
2848 | mlir::Value val = fir::getBase(converter.genExprValue(arg, naCtx, &loc)); |
2849 | overrides.try_emplace(&arg, val); |
2850 | } |
2851 | } |
2852 | |
2853 | builder.restoreInsertionPoint(atomicAt); |
2854 | auto updateOp = |
2855 | builder.create<mlir::omp::AtomicUpdateOp>(loc, atomAddr, hint, memOrder); |
2856 | |
2857 | mlir::Region ®ion = updateOp->getRegion(0); |
2858 | mlir::Block *block = builder.createBlock(®ion, {}, {atomType}, {loc}); |
2859 | mlir::Value localAtom = fir::getBase(block->getArgument(0)); |
2860 | overrides.try_emplace(&atom, localAtom); |
2861 | |
2862 | converter.overrideExprValues(&overrides); |
2863 | mlir::Value updated = |
2864 | fir::getBase(converter.genExprValue(assign.rhs, stmtCtx, &loc)); |
2865 | mlir::Value converted = builder.createConvert(loc, atomType, updated); |
2866 | builder.create<mlir::omp::YieldOp>(loc, converted); |
2867 | converter.resetExprOverrides(); |
2868 | |
2869 | builder.restoreInsertionPoint(postAt); // For naCtx cleanups |
2870 | return updateOp; |
2871 | } |
2872 | |
2873 | static mlir::Operation * |
2874 | genAtomicOperation(lower::AbstractConverter &converter, mlir::Location loc, |
2875 | lower::StatementContext &stmtCtx, int action, |
2876 | mlir::Value atomAddr, const semantics::SomeExpr &atom, |
2877 | const evaluate::Assignment &assign, mlir::IntegerAttr hint, |
2878 | mlir::omp::ClauseMemoryOrderKindAttr memOrder, |
2879 | fir::FirOpBuilder::InsertPoint preAt, |
2880 | fir::FirOpBuilder::InsertPoint atomicAt, |
2881 | fir::FirOpBuilder::InsertPoint postAt) { |
2882 | if (isPointerAssignment(assign)) { |
2883 | TODO(loc, "Code generation for pointer assignment is not implemented yet" ); |
2884 | } |
2885 | |
2886 | // This function and the functions called here do not preserve the |
2887 | // builder's insertion point, or set it to anything specific. |
2888 | switch (action) { |
2889 | case parser::OpenMPAtomicConstruct::Analysis::Read: |
2890 | return genAtomicRead(converter, loc, stmtCtx, atomAddr, atom, assign, hint, |
2891 | memOrder, preAt, atomicAt, postAt); |
2892 | case parser::OpenMPAtomicConstruct::Analysis::Write: |
2893 | return genAtomicWrite(converter, loc, stmtCtx, atomAddr, atom, assign, hint, |
2894 | memOrder, preAt, atomicAt, postAt); |
2895 | case parser::OpenMPAtomicConstruct::Analysis::Update: |
2896 | return genAtomicUpdate(converter, loc, stmtCtx, atomAddr, atom, assign, |
2897 | hint, memOrder, preAt, atomicAt, postAt); |
2898 | default: |
2899 | return nullptr; |
2900 | } |
2901 | } |
2902 | |
2903 | //===----------------------------------------------------------------------===// |
2904 | // Code generation functions for the standalone version of constructs that can |
2905 | // also be a leaf of a composite construct |
2906 | //===----------------------------------------------------------------------===// |
2907 | |
2908 | static mlir::omp::DistributeOp genStandaloneDistribute( |
2909 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2910 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2911 | lower::pft::Evaluation &eval, mlir::Location loc, |
2912 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2913 | mlir::omp::DistributeOperands distributeClauseOps; |
2914 | genDistributeClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
2915 | distributeClauseOps); |
2916 | |
2917 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2918 | /*shouldCollectPreDeterminedSymbols=*/true, |
2919 | enableDelayedPrivatization, symTable); |
2920 | dsp.processStep1(clauseOps: &distributeClauseOps); |
2921 | |
2922 | mlir::omp::LoopNestOperands loopNestClauseOps; |
2923 | llvm::SmallVector<const semantics::Symbol *> iv; |
2924 | genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc, |
2925 | loopNestClauseOps, iv); |
2926 | |
2927 | EntryBlockArgs distributeArgs; |
2928 | distributeArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
2929 | distributeArgs.priv.vars = distributeClauseOps.privateVars; |
2930 | auto distributeOp = genWrapperOp<mlir::omp::DistributeOp>( |
2931 | converter, loc, distributeClauseOps, distributeArgs); |
2932 | |
2933 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item, |
2934 | loopNestClauseOps, iv, {{distributeOp, distributeArgs}}, |
2935 | llvm::omp::Directive::OMPD_distribute, dsp); |
2936 | return distributeOp; |
2937 | } |
2938 | |
2939 | static mlir::omp::WsloopOp genStandaloneDo( |
2940 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2941 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2942 | lower::pft::Evaluation &eval, mlir::Location loc, |
2943 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2944 | mlir::omp::WsloopOperands wsloopClauseOps; |
2945 | llvm::SmallVector<const semantics::Symbol *> wsloopReductionSyms; |
2946 | genWsloopClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
2947 | wsloopClauseOps, wsloopReductionSyms); |
2948 | |
2949 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
2950 | /*shouldCollectPreDeterminedSymbols=*/true, |
2951 | enableDelayedPrivatization, symTable); |
2952 | dsp.processStep1(clauseOps: &wsloopClauseOps); |
2953 | |
2954 | mlir::omp::LoopNestOperands loopNestClauseOps; |
2955 | llvm::SmallVector<const semantics::Symbol *> iv; |
2956 | genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc, |
2957 | loopNestClauseOps, iv); |
2958 | |
2959 | EntryBlockArgs wsloopArgs; |
2960 | wsloopArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
2961 | wsloopArgs.priv.vars = wsloopClauseOps.privateVars; |
2962 | wsloopArgs.reduction.syms = wsloopReductionSyms; |
2963 | wsloopArgs.reduction.vars = wsloopClauseOps.reductionVars; |
2964 | auto wsloopOp = genWrapperOp<mlir::omp::WsloopOp>( |
2965 | converter, loc, wsloopClauseOps, wsloopArgs); |
2966 | |
2967 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item, |
2968 | loopNestClauseOps, iv, {{wsloopOp, wsloopArgs}}, |
2969 | llvm::omp::Directive::OMPD_do, dsp); |
2970 | return wsloopOp; |
2971 | } |
2972 | |
2973 | static mlir::omp::ParallelOp genStandaloneParallel( |
2974 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
2975 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
2976 | lower::pft::Evaluation &eval, mlir::Location loc, |
2977 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
2978 | mlir::omp::ParallelOperands parallelClauseOps; |
2979 | llvm::SmallVector<const semantics::Symbol *> parallelReductionSyms; |
2980 | genParallelClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
2981 | parallelClauseOps, parallelReductionSyms); |
2982 | |
2983 | std::optional<DataSharingProcessor> dsp; |
2984 | if (enableDelayedPrivatization) { |
2985 | dsp.emplace(converter, semaCtx, item->clauses, eval, |
2986 | lower::omp::isLastItemInQueue(item, queue), |
2987 | /*useDelayedPrivatization=*/true, symTable); |
2988 | dsp->processStep1(clauseOps: ¶llelClauseOps); |
2989 | } |
2990 | |
2991 | EntryBlockArgs parallelArgs; |
2992 | if (dsp) |
2993 | parallelArgs.priv.syms = dsp->getDelayedPrivSymbols(); |
2994 | parallelArgs.priv.vars = parallelClauseOps.privateVars; |
2995 | parallelArgs.reduction.syms = parallelReductionSyms; |
2996 | parallelArgs.reduction.vars = parallelClauseOps.reductionVars; |
2997 | return genParallelOp(converter, symTable, semaCtx, eval, loc, queue, item, |
2998 | parallelClauseOps, parallelArgs, |
2999 | enableDelayedPrivatization ? &dsp.value() : nullptr); |
3000 | } |
3001 | |
3002 | static mlir::omp::SimdOp |
3003 | genStandaloneSimd(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3004 | semantics::SemanticsContext &semaCtx, |
3005 | lower::pft::Evaluation &eval, mlir::Location loc, |
3006 | const ConstructQueue &queue, |
3007 | ConstructQueue::const_iterator item) { |
3008 | mlir::omp::SimdOperands simdClauseOps; |
3009 | llvm::SmallVector<const semantics::Symbol *> simdReductionSyms; |
3010 | genSimdClauses(converter, semaCtx, item->clauses, loc, simdClauseOps, |
3011 | simdReductionSyms); |
3012 | |
3013 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
3014 | /*shouldCollectPreDeterminedSymbols=*/true, |
3015 | enableDelayedPrivatization, symTable); |
3016 | dsp.processStep1(clauseOps: &simdClauseOps); |
3017 | |
3018 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3019 | llvm::SmallVector<const semantics::Symbol *> iv; |
3020 | genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc, |
3021 | loopNestClauseOps, iv); |
3022 | |
3023 | EntryBlockArgs simdArgs; |
3024 | simdArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
3025 | simdArgs.priv.vars = simdClauseOps.privateVars; |
3026 | simdArgs.reduction.syms = simdReductionSyms; |
3027 | simdArgs.reduction.vars = simdClauseOps.reductionVars; |
3028 | auto simdOp = |
3029 | genWrapperOp<mlir::omp::SimdOp>(converter, loc, simdClauseOps, simdArgs); |
3030 | |
3031 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item, |
3032 | loopNestClauseOps, iv, {{simdOp, simdArgs}}, |
3033 | llvm::omp::Directive::OMPD_simd, dsp); |
3034 | return simdOp; |
3035 | } |
3036 | |
3037 | static mlir::omp::TaskloopOp genStandaloneTaskloop( |
3038 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3039 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3040 | lower::pft::Evaluation &eval, mlir::Location loc, |
3041 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3042 | mlir::omp::TaskloopOperands taskloopClauseOps; |
3043 | genTaskloopClauses(converter, semaCtx, stmtCtx, item->clauses, loc, |
3044 | taskloopClauseOps); |
3045 | DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval, |
3046 | /*shouldCollectPreDeterminedSymbols=*/true, |
3047 | enableDelayedPrivatization, symTable); |
3048 | dsp.processStep1(clauseOps: &taskloopClauseOps); |
3049 | |
3050 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3051 | llvm::SmallVector<const semantics::Symbol *> iv; |
3052 | genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc, |
3053 | loopNestClauseOps, iv); |
3054 | |
3055 | EntryBlockArgs taskloopArgs; |
3056 | taskloopArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
3057 | taskloopArgs.priv.vars = taskloopClauseOps.privateVars; |
3058 | |
3059 | auto taskLoopOp = genWrapperOp<mlir::omp::TaskloopOp>( |
3060 | converter, loc, taskloopClauseOps, taskloopArgs); |
3061 | |
3062 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item, |
3063 | loopNestClauseOps, iv, {{taskLoopOp, taskloopArgs}}, |
3064 | llvm::omp::Directive::OMPD_taskloop, dsp); |
3065 | return taskLoopOp; |
3066 | } |
3067 | |
3068 | //===----------------------------------------------------------------------===// |
3069 | // Code generation functions for composite constructs |
3070 | //===----------------------------------------------------------------------===// |
3071 | |
3072 | static mlir::omp::DistributeOp genCompositeDistributeParallelDo( |
3073 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3074 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3075 | lower::pft::Evaluation &eval, mlir::Location loc, |
3076 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3077 | assert(std::distance(item, queue.end()) == 3 && "Invalid leaf constructs" ); |
3078 | ConstructQueue::const_iterator distributeItem = item; |
3079 | ConstructQueue::const_iterator parallelItem = std::next(x: distributeItem); |
3080 | ConstructQueue::const_iterator doItem = std::next(x: parallelItem); |
3081 | |
3082 | // Create parent omp.parallel first. |
3083 | mlir::omp::ParallelOperands parallelClauseOps; |
3084 | llvm::SmallVector<const semantics::Symbol *> parallelReductionSyms; |
3085 | genParallelClauses(converter, semaCtx, stmtCtx, parallelItem->clauses, loc, |
3086 | parallelClauseOps, parallelReductionSyms); |
3087 | |
3088 | DataSharingProcessor dsp(converter, semaCtx, doItem->clauses, eval, |
3089 | /*shouldCollectPreDeterminedSymbols=*/true, |
3090 | /*useDelayedPrivatization=*/true, symTable); |
3091 | dsp.processStep1(clauseOps: ¶llelClauseOps); |
3092 | |
3093 | EntryBlockArgs parallelArgs; |
3094 | parallelArgs.priv.syms = dsp.getDelayedPrivSymbols(); |
3095 | parallelArgs.priv.vars = parallelClauseOps.privateVars; |
3096 | parallelArgs.reduction.syms = parallelReductionSyms; |
3097 | parallelArgs.reduction.vars = parallelClauseOps.reductionVars; |
3098 | genParallelOp(converter, symTable, semaCtx, eval, loc, queue, parallelItem, |
3099 | parallelClauseOps, parallelArgs, &dsp, /*isComposite=*/true); |
3100 | |
3101 | // Clause processing. |
3102 | mlir::omp::DistributeOperands distributeClauseOps; |
3103 | genDistributeClauses(converter, semaCtx, stmtCtx, distributeItem->clauses, |
3104 | loc, distributeClauseOps); |
3105 | |
3106 | mlir::omp::WsloopOperands wsloopClauseOps; |
3107 | llvm::SmallVector<const semantics::Symbol *> wsloopReductionSyms; |
3108 | genWsloopClauses(converter, semaCtx, stmtCtx, doItem->clauses, loc, |
3109 | wsloopClauseOps, wsloopReductionSyms); |
3110 | |
3111 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3112 | llvm::SmallVector<const semantics::Symbol *> iv; |
3113 | genLoopNestClauses(converter, semaCtx, eval, doItem->clauses, loc, |
3114 | loopNestClauseOps, iv); |
3115 | |
3116 | // Operation creation. |
3117 | EntryBlockArgs distributeArgs; |
3118 | // TODO: Add private syms and vars. |
3119 | auto distributeOp = genWrapperOp<mlir::omp::DistributeOp>( |
3120 | converter, loc, distributeClauseOps, distributeArgs); |
3121 | distributeOp.setComposite(/*val=*/true); |
3122 | |
3123 | EntryBlockArgs wsloopArgs; |
3124 | // TODO: Add private syms and vars. |
3125 | wsloopArgs.reduction.syms = wsloopReductionSyms; |
3126 | wsloopArgs.reduction.vars = wsloopClauseOps.reductionVars; |
3127 | auto wsloopOp = genWrapperOp<mlir::omp::WsloopOp>( |
3128 | converter, loc, wsloopClauseOps, wsloopArgs); |
3129 | wsloopOp.setComposite(/*val=*/true); |
3130 | |
3131 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, doItem, |
3132 | loopNestClauseOps, iv, |
3133 | {{distributeOp, distributeArgs}, {wsloopOp, wsloopArgs}}, |
3134 | llvm::omp::Directive::OMPD_distribute_parallel_do, dsp); |
3135 | return distributeOp; |
3136 | } |
3137 | |
3138 | static mlir::omp::DistributeOp genCompositeDistributeParallelDoSimd( |
3139 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3140 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3141 | lower::pft::Evaluation &eval, mlir::Location loc, |
3142 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3143 | assert(std::distance(item, queue.end()) == 4 && "Invalid leaf constructs" ); |
3144 | ConstructQueue::const_iterator distributeItem = item; |
3145 | ConstructQueue::const_iterator parallelItem = std::next(x: distributeItem); |
3146 | ConstructQueue::const_iterator doItem = std::next(x: parallelItem); |
3147 | ConstructQueue::const_iterator simdItem = std::next(x: doItem); |
3148 | |
3149 | // Create parent omp.parallel first. |
3150 | mlir::omp::ParallelOperands parallelClauseOps; |
3151 | llvm::SmallVector<const semantics::Symbol *> parallelReductionSyms; |
3152 | genParallelClauses(converter, semaCtx, stmtCtx, parallelItem->clauses, loc, |
3153 | parallelClauseOps, parallelReductionSyms); |
3154 | |
3155 | DataSharingProcessor parallelItemDSP( |
3156 | converter, semaCtx, parallelItem->clauses, eval, |
3157 | /*shouldCollectPreDeterminedSymbols=*/false, |
3158 | /*useDelayedPrivatization=*/true, symTable); |
3159 | parallelItemDSP.processStep1(clauseOps: ¶llelClauseOps); |
3160 | |
3161 | EntryBlockArgs parallelArgs; |
3162 | parallelArgs.priv.syms = parallelItemDSP.getDelayedPrivSymbols(); |
3163 | parallelArgs.priv.vars = parallelClauseOps.privateVars; |
3164 | parallelArgs.reduction.syms = parallelReductionSyms; |
3165 | parallelArgs.reduction.vars = parallelClauseOps.reductionVars; |
3166 | genParallelOp(converter, symTable, semaCtx, eval, loc, queue, parallelItem, |
3167 | parallelClauseOps, parallelArgs, ¶llelItemDSP, |
3168 | /*isComposite=*/true); |
3169 | |
3170 | // Clause processing. |
3171 | mlir::omp::DistributeOperands distributeClauseOps; |
3172 | genDistributeClauses(converter, semaCtx, stmtCtx, distributeItem->clauses, |
3173 | loc, distributeClauseOps); |
3174 | |
3175 | mlir::omp::WsloopOperands wsloopClauseOps; |
3176 | llvm::SmallVector<const semantics::Symbol *> wsloopReductionSyms; |
3177 | genWsloopClauses(converter, semaCtx, stmtCtx, doItem->clauses, loc, |
3178 | wsloopClauseOps, wsloopReductionSyms); |
3179 | |
3180 | mlir::omp::SimdOperands simdClauseOps; |
3181 | llvm::SmallVector<const semantics::Symbol *> simdReductionSyms; |
3182 | genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps, |
3183 | simdReductionSyms); |
3184 | |
3185 | DataSharingProcessor simdItemDSP(converter, semaCtx, simdItem->clauses, eval, |
3186 | /*shouldCollectPreDeterminedSymbols=*/true, |
3187 | /*useDelayedPrivatization=*/true, symTable); |
3188 | simdItemDSP.processStep1(clauseOps: &simdClauseOps); |
3189 | |
3190 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3191 | llvm::SmallVector<const semantics::Symbol *> iv; |
3192 | genLoopNestClauses(converter, semaCtx, eval, simdItem->clauses, loc, |
3193 | loopNestClauseOps, iv); |
3194 | |
3195 | // Operation creation. |
3196 | EntryBlockArgs distributeArgs; |
3197 | // TODO: Add private syms and vars. |
3198 | auto distributeOp = genWrapperOp<mlir::omp::DistributeOp>( |
3199 | converter, loc, distributeClauseOps, distributeArgs); |
3200 | distributeOp.setComposite(/*val=*/true); |
3201 | |
3202 | EntryBlockArgs wsloopArgs; |
3203 | // TODO: Add private syms and vars. |
3204 | wsloopArgs.reduction.syms = wsloopReductionSyms; |
3205 | wsloopArgs.reduction.vars = wsloopClauseOps.reductionVars; |
3206 | auto wsloopOp = genWrapperOp<mlir::omp::WsloopOp>( |
3207 | converter, loc, wsloopClauseOps, wsloopArgs); |
3208 | wsloopOp.setComposite(/*val=*/true); |
3209 | |
3210 | EntryBlockArgs simdArgs; |
3211 | simdArgs.priv.syms = simdItemDSP.getDelayedPrivSymbols(); |
3212 | simdArgs.priv.vars = simdClauseOps.privateVars; |
3213 | simdArgs.reduction.syms = simdReductionSyms; |
3214 | simdArgs.reduction.vars = simdClauseOps.reductionVars; |
3215 | auto simdOp = |
3216 | genWrapperOp<mlir::omp::SimdOp>(converter, loc, simdClauseOps, simdArgs); |
3217 | simdOp.setComposite(/*val=*/true); |
3218 | |
3219 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, simdItem, |
3220 | loopNestClauseOps, iv, |
3221 | {{distributeOp, distributeArgs}, |
3222 | {wsloopOp, wsloopArgs}, |
3223 | {simdOp, simdArgs}}, |
3224 | llvm::omp::Directive::OMPD_distribute_parallel_do_simd, |
3225 | simdItemDSP); |
3226 | return distributeOp; |
3227 | } |
3228 | |
3229 | static mlir::omp::DistributeOp genCompositeDistributeSimd( |
3230 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3231 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3232 | lower::pft::Evaluation &eval, mlir::Location loc, |
3233 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3234 | assert(std::distance(item, queue.end()) == 2 && "Invalid leaf constructs" ); |
3235 | ConstructQueue::const_iterator distributeItem = item; |
3236 | ConstructQueue::const_iterator simdItem = std::next(x: distributeItem); |
3237 | |
3238 | // Clause processing. |
3239 | mlir::omp::DistributeOperands distributeClauseOps; |
3240 | genDistributeClauses(converter, semaCtx, stmtCtx, distributeItem->clauses, |
3241 | loc, distributeClauseOps); |
3242 | |
3243 | mlir::omp::SimdOperands simdClauseOps; |
3244 | llvm::SmallVector<const semantics::Symbol *> simdReductionSyms; |
3245 | genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps, |
3246 | simdReductionSyms); |
3247 | |
3248 | // TODO: Support delayed privatization. |
3249 | DataSharingProcessor dsp(converter, semaCtx, simdItem->clauses, eval, |
3250 | /*shouldCollectPreDeterminedSymbols=*/true, |
3251 | /*useDelayedPrivatization=*/false, symTable); |
3252 | dsp.processStep1(); |
3253 | |
3254 | // Pass the innermost leaf construct's clauses because that's where COLLAPSE |
3255 | // is placed by construct decomposition. |
3256 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3257 | llvm::SmallVector<const semantics::Symbol *> iv; |
3258 | genLoopNestClauses(converter, semaCtx, eval, simdItem->clauses, loc, |
3259 | loopNestClauseOps, iv); |
3260 | |
3261 | // Operation creation. |
3262 | EntryBlockArgs distributeArgs; |
3263 | // TODO: Add private syms and vars. |
3264 | auto distributeOp = genWrapperOp<mlir::omp::DistributeOp>( |
3265 | converter, loc, distributeClauseOps, distributeArgs); |
3266 | distributeOp.setComposite(/*val=*/true); |
3267 | |
3268 | EntryBlockArgs simdArgs; |
3269 | // TODO: Add private syms and vars. |
3270 | simdArgs.reduction.syms = simdReductionSyms; |
3271 | simdArgs.reduction.vars = simdClauseOps.reductionVars; |
3272 | auto simdOp = |
3273 | genWrapperOp<mlir::omp::SimdOp>(converter, loc, simdClauseOps, simdArgs); |
3274 | simdOp.setComposite(/*val=*/true); |
3275 | |
3276 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, simdItem, |
3277 | loopNestClauseOps, iv, |
3278 | {{distributeOp, distributeArgs}, {simdOp, simdArgs}}, |
3279 | llvm::omp::Directive::OMPD_distribute_simd, dsp); |
3280 | return distributeOp; |
3281 | } |
3282 | |
3283 | static mlir::omp::WsloopOp genCompositeDoSimd( |
3284 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3285 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3286 | lower::pft::Evaluation &eval, mlir::Location loc, |
3287 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3288 | assert(std::distance(item, queue.end()) == 2 && "Invalid leaf constructs" ); |
3289 | ConstructQueue::const_iterator doItem = item; |
3290 | ConstructQueue::const_iterator simdItem = std::next(x: doItem); |
3291 | |
3292 | // Clause processing. |
3293 | mlir::omp::WsloopOperands wsloopClauseOps; |
3294 | llvm::SmallVector<const semantics::Symbol *> wsloopReductionSyms; |
3295 | genWsloopClauses(converter, semaCtx, stmtCtx, doItem->clauses, loc, |
3296 | wsloopClauseOps, wsloopReductionSyms); |
3297 | |
3298 | mlir::omp::SimdOperands simdClauseOps; |
3299 | llvm::SmallVector<const semantics::Symbol *> simdReductionSyms; |
3300 | genSimdClauses(converter, semaCtx, simdItem->clauses, loc, simdClauseOps, |
3301 | simdReductionSyms); |
3302 | |
3303 | // TODO: Support delayed privatization. |
3304 | DataSharingProcessor dsp(converter, semaCtx, simdItem->clauses, eval, |
3305 | /*shouldCollectPreDeterminedSymbols=*/true, |
3306 | /*useDelayedPrivatization=*/false, symTable); |
3307 | dsp.processStep1(); |
3308 | |
3309 | // Pass the innermost leaf construct's clauses because that's where COLLAPSE |
3310 | // is placed by construct decomposition. |
3311 | mlir::omp::LoopNestOperands loopNestClauseOps; |
3312 | llvm::SmallVector<const semantics::Symbol *> iv; |
3313 | genLoopNestClauses(converter, semaCtx, eval, simdItem->clauses, loc, |
3314 | loopNestClauseOps, iv); |
3315 | |
3316 | // Operation creation. |
3317 | EntryBlockArgs wsloopArgs; |
3318 | // TODO: Add private syms and vars. |
3319 | wsloopArgs.reduction.syms = wsloopReductionSyms; |
3320 | wsloopArgs.reduction.vars = wsloopClauseOps.reductionVars; |
3321 | auto wsloopOp = genWrapperOp<mlir::omp::WsloopOp>( |
3322 | converter, loc, wsloopClauseOps, wsloopArgs); |
3323 | wsloopOp.setComposite(/*val=*/true); |
3324 | |
3325 | EntryBlockArgs simdArgs; |
3326 | // TODO: Add private syms and vars. |
3327 | simdArgs.reduction.syms = simdReductionSyms; |
3328 | simdArgs.reduction.vars = simdClauseOps.reductionVars; |
3329 | auto simdOp = |
3330 | genWrapperOp<mlir::omp::SimdOp>(converter, loc, simdClauseOps, simdArgs); |
3331 | simdOp.setComposite(/*val=*/true); |
3332 | |
3333 | genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, simdItem, |
3334 | loopNestClauseOps, iv, |
3335 | {{wsloopOp, wsloopArgs}, {simdOp, simdArgs}}, |
3336 | llvm::omp::Directive::OMPD_do_simd, dsp); |
3337 | return wsloopOp; |
3338 | } |
3339 | |
3340 | static mlir::omp::TaskloopOp genCompositeTaskloopSimd( |
3341 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3342 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3343 | lower::pft::Evaluation &eval, mlir::Location loc, |
3344 | const ConstructQueue &queue, ConstructQueue::const_iterator item) { |
3345 | assert(std::distance(item, queue.end()) == 2 && "Invalid leaf constructs" ); |
3346 | TODO(loc, "Composite TASKLOOP SIMD" ); |
3347 | return nullptr; |
3348 | } |
3349 | |
3350 | //===----------------------------------------------------------------------===// |
3351 | // Dispatch |
3352 | //===----------------------------------------------------------------------===// |
3353 | |
3354 | static bool genOMPCompositeDispatch( |
3355 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3356 | lower::StatementContext &stmtCtx, semantics::SemanticsContext &semaCtx, |
3357 | lower::pft::Evaluation &eval, mlir::Location loc, |
3358 | const ConstructQueue &queue, ConstructQueue::const_iterator item, |
3359 | mlir::Operation *&newOp) { |
3360 | using llvm::omp::Directive; |
3361 | using lower::omp::matchLeafSequence; |
3362 | |
3363 | // TODO: Privatization for composite constructs is currently only done based |
3364 | // on the clauses for their last leaf construct, which may not always be |
3365 | // correct. Consider per-leaf privatization of composite constructs once |
3366 | // delayed privatization is supported by all participating ops. |
3367 | if (matchLeafSequence(item, queue, Directive::OMPD_distribute_parallel_do)) |
3368 | newOp = genCompositeDistributeParallelDo(converter, symTable, stmtCtx, |
3369 | semaCtx, eval, loc, queue, item); |
3370 | else if (matchLeafSequence(item, queue, |
3371 | Directive::OMPD_distribute_parallel_do_simd)) |
3372 | newOp = genCompositeDistributeParallelDoSimd( |
3373 | converter, symTable, stmtCtx, semaCtx, eval, loc, queue, item); |
3374 | else if (matchLeafSequence(item, queue, Directive::OMPD_distribute_simd)) |
3375 | newOp = genCompositeDistributeSimd(converter, symTable, stmtCtx, semaCtx, |
3376 | eval, loc, queue, item); |
3377 | else if (matchLeafSequence(item, queue, Directive::OMPD_do_simd)) |
3378 | newOp = genCompositeDoSimd(converter, symTable, stmtCtx, semaCtx, eval, loc, |
3379 | queue, item); |
3380 | else if (matchLeafSequence(item, queue, Directive::OMPD_taskloop_simd)) |
3381 | newOp = genCompositeTaskloopSimd(converter, symTable, stmtCtx, semaCtx, |
3382 | eval, loc, queue, item); |
3383 | else |
3384 | return false; |
3385 | |
3386 | return true; |
3387 | } |
3388 | |
3389 | static void genOMPDispatch(lower::AbstractConverter &converter, |
3390 | lower::SymMap &symTable, |
3391 | semantics::SemanticsContext &semaCtx, |
3392 | lower::pft::Evaluation &eval, mlir::Location loc, |
3393 | const ConstructQueue &queue, |
3394 | ConstructQueue::const_iterator item) { |
3395 | assert(item != queue.end()); |
3396 | |
3397 | lower::StatementContext stmtCtx; |
3398 | mlir::Operation *newOp = nullptr; |
3399 | |
3400 | // Generate cleanup code for the stmtCtx after newOp |
3401 | auto finalizeStmtCtx = [&]() { |
3402 | if (newOp) { |
3403 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
3404 | fir::FirOpBuilder::InsertionGuard guard(builder); |
3405 | builder.setInsertionPointAfter(newOp); |
3406 | stmtCtx.finalizeAndPop(); |
3407 | } |
3408 | }; |
3409 | |
3410 | bool loopLeaf = llvm::omp::getDirectiveAssociation(item->id) == |
3411 | llvm::omp::Association::Loop; |
3412 | if (loopLeaf) { |
3413 | symTable.pushScope(); |
3414 | if (genOMPCompositeDispatch(converter, symTable, stmtCtx, semaCtx, eval, |
3415 | loc, queue, item, newOp)) { |
3416 | symTable.popScope(); |
3417 | finalizeStmtCtx(); |
3418 | return; |
3419 | } |
3420 | } |
3421 | |
3422 | switch (llvm::omp::Directive dir = item->id) { |
3423 | case llvm::omp::Directive::OMPD_barrier: |
3424 | newOp = genBarrierOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3425 | break; |
3426 | case llvm::omp::Directive::OMPD_distribute: |
3427 | newOp = genStandaloneDistribute(converter, symTable, stmtCtx, semaCtx, eval, |
3428 | loc, queue, item); |
3429 | break; |
3430 | case llvm::omp::Directive::OMPD_do: |
3431 | newOp = genStandaloneDo(converter, symTable, stmtCtx, semaCtx, eval, loc, |
3432 | queue, item); |
3433 | break; |
3434 | case llvm::omp::Directive::OMPD_loop: |
3435 | newOp = genLoopOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3436 | break; |
3437 | case llvm::omp::Directive::OMPD_masked: |
3438 | newOp = genMaskedOp(converter, symTable, stmtCtx, semaCtx, eval, loc, queue, |
3439 | item); |
3440 | break; |
3441 | case llvm::omp::Directive::OMPD_master: |
3442 | newOp = genMasterOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3443 | break; |
3444 | case llvm::omp::Directive::OMPD_ordered: |
3445 | // Block-associated "ordered" construct. |
3446 | newOp = genOrderedRegionOp(converter, symTable, semaCtx, eval, loc, queue, |
3447 | item); |
3448 | break; |
3449 | case llvm::omp::Directive::OMPD_parallel: |
3450 | newOp = genStandaloneParallel(converter, symTable, stmtCtx, semaCtx, eval, |
3451 | loc, queue, item); |
3452 | break; |
3453 | case llvm::omp::Directive::OMPD_scan: |
3454 | newOp = genScanOp(converter, symTable, semaCtx, loc, queue, item); |
3455 | break; |
3456 | case llvm::omp::Directive::OMPD_section: |
3457 | llvm_unreachable("genOMPDispatch: OMPD_section" ); |
3458 | // Lowered in the enclosing genSectionsOp. |
3459 | break; |
3460 | case llvm::omp::Directive::OMPD_sections: |
3461 | // Called directly from genOMP([...], OpenMPSectionsConstruct) because it |
3462 | // has a different prototype. |
3463 | // This code path is still taken when iterating through the construct queue |
3464 | // in genBodyOfOp |
3465 | break; |
3466 | case llvm::omp::Directive::OMPD_simd: |
3467 | newOp = |
3468 | genStandaloneSimd(converter, symTable, semaCtx, eval, loc, queue, item); |
3469 | break; |
3470 | case llvm::omp::Directive::OMPD_scope: |
3471 | newOp = genScopeOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3472 | break; |
3473 | case llvm::omp::Directive::OMPD_single: |
3474 | newOp = genSingleOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3475 | break; |
3476 | case llvm::omp::Directive::OMPD_target: |
3477 | newOp = genTargetOp(converter, symTable, stmtCtx, semaCtx, eval, loc, queue, |
3478 | item); |
3479 | break; |
3480 | case llvm::omp::Directive::OMPD_target_data: |
3481 | newOp = genTargetDataOp(converter, symTable, stmtCtx, semaCtx, eval, loc, |
3482 | queue, item); |
3483 | break; |
3484 | case llvm::omp::Directive::OMPD_target_enter_data: |
3485 | newOp = genTargetEnterExitUpdateDataOp<mlir::omp::TargetEnterDataOp>( |
3486 | converter, symTable, stmtCtx, semaCtx, loc, queue, item); |
3487 | break; |
3488 | case llvm::omp::Directive::OMPD_target_exit_data: |
3489 | newOp = genTargetEnterExitUpdateDataOp<mlir::omp::TargetExitDataOp>( |
3490 | converter, symTable, stmtCtx, semaCtx, loc, queue, item); |
3491 | break; |
3492 | case llvm::omp::Directive::OMPD_target_update: |
3493 | newOp = genTargetEnterExitUpdateDataOp<mlir::omp::TargetUpdateOp>( |
3494 | converter, symTable, stmtCtx, semaCtx, loc, queue, item); |
3495 | break; |
3496 | case llvm::omp::Directive::OMPD_task: |
3497 | newOp = genTaskOp(converter, symTable, stmtCtx, semaCtx, eval, loc, queue, |
3498 | item); |
3499 | break; |
3500 | case llvm::omp::Directive::OMPD_taskgroup: |
3501 | newOp = |
3502 | genTaskgroupOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3503 | break; |
3504 | case llvm::omp::Directive::OMPD_taskloop: |
3505 | newOp = genStandaloneTaskloop(converter, symTable, stmtCtx, semaCtx, eval, |
3506 | loc, queue, item); |
3507 | break; |
3508 | case llvm::omp::Directive::OMPD_taskwait: |
3509 | newOp = genTaskwaitOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3510 | break; |
3511 | case llvm::omp::Directive::OMPD_taskyield: |
3512 | newOp = |
3513 | genTaskyieldOp(converter, symTable, semaCtx, eval, loc, queue, item); |
3514 | break; |
3515 | case llvm::omp::Directive::OMPD_teams: |
3516 | newOp = genTeamsOp(converter, symTable, stmtCtx, semaCtx, eval, loc, queue, |
3517 | item); |
3518 | break; |
3519 | case llvm::omp::Directive::OMPD_tile: |
3520 | case llvm::omp::Directive::OMPD_unroll: { |
3521 | unsigned version = semaCtx.langOptions().OpenMPVersion; |
3522 | TODO(loc, "Unhandled loop directive (" + |
3523 | llvm::omp::getOpenMPDirectiveName(dir, version) + ")" ); |
3524 | } |
3525 | // case llvm::omp::Directive::OMPD_workdistribute: |
3526 | case llvm::omp::Directive::OMPD_workshare: |
3527 | newOp = genWorkshareOp(converter, symTable, stmtCtx, semaCtx, eval, loc, |
3528 | queue, item); |
3529 | break; |
3530 | default: |
3531 | // Combined and composite constructs should have been split into a sequence |
3532 | // of leaf constructs when building the construct queue. |
3533 | assert(!llvm::omp::isLeafConstruct(dir) && |
3534 | "Unexpected compound construct." ); |
3535 | break; |
3536 | } |
3537 | |
3538 | finalizeStmtCtx(); |
3539 | if (loopLeaf) |
3540 | symTable.popScope(); |
3541 | } |
3542 | |
3543 | //===----------------------------------------------------------------------===// |
3544 | // OpenMPDeclarativeConstruct visitors |
3545 | //===----------------------------------------------------------------------===// |
3546 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3547 | semantics::SemanticsContext &semaCtx, |
3548 | lower::pft::Evaluation &eval, |
3549 | const parser::OpenMPUtilityConstruct &); |
3550 | |
3551 | static void |
3552 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3553 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3554 | const parser::OpenMPDeclarativeAllocate &declarativeAllocate) { |
3555 | TODO(converter.getCurrentLocation(), "OpenMPDeclarativeAllocate" ); |
3556 | } |
3557 | |
3558 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3559 | semantics::SemanticsContext &semaCtx, |
3560 | lower::pft::Evaluation &eval, |
3561 | const parser::OpenMPDeclarativeAssumes &assumesConstruct) { |
3562 | TODO(converter.getCurrentLocation(), "OpenMP ASSUMES declaration" ); |
3563 | } |
3564 | |
3565 | static void |
3566 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3567 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3568 | const parser::OmpDeclareVariantDirective &declareVariantDirective) { |
3569 | TODO(converter.getCurrentLocation(), "OmpDeclareVariantDirective" ); |
3570 | } |
3571 | |
3572 | static void genOMP( |
3573 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
3574 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3575 | const parser::OpenMPDeclareReductionConstruct &declareReductionConstruct) { |
3576 | TODO(converter.getCurrentLocation(), "OpenMPDeclareReductionConstruct" ); |
3577 | } |
3578 | |
3579 | static void |
3580 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3581 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3582 | const parser::OpenMPDeclareSimdConstruct &declareSimdConstruct) { |
3583 | TODO(converter.getCurrentLocation(), "OpenMPDeclareSimdConstruct" ); |
3584 | } |
3585 | |
3586 | static void |
3587 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3588 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3589 | const parser::OpenMPDeclareMapperConstruct &declareMapperConstruct) { |
3590 | mlir::Location loc = converter.genLocation(declareMapperConstruct.source); |
3591 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
3592 | lower::StatementContext stmtCtx; |
3593 | const auto &spec = |
3594 | std::get<parser::OmpMapperSpecifier>(declareMapperConstruct.t); |
3595 | const auto &mapperName{std::get<std::string>(spec.t)}; |
3596 | const auto &varType{std::get<parser::TypeSpec>(spec.t)}; |
3597 | const auto &varName{std::get<parser::Name>(spec.t)}; |
3598 | assert(varType.declTypeSpec->category() == |
3599 | semantics::DeclTypeSpec::Category::TypeDerived && |
3600 | "Expected derived type" ); |
3601 | |
3602 | std::string mapperNameStr = mapperName; |
3603 | if (auto *sym = converter.getCurrentScope().FindSymbol(mapperNameStr)) |
3604 | mapperNameStr = converter.mangleName(mapperNameStr, sym->owner()); |
3605 | |
3606 | // Save current insertion point before moving to the module scope to create |
3607 | // the DeclareMapperOp |
3608 | mlir::OpBuilder::InsertionGuard guard(firOpBuilder); |
3609 | |
3610 | firOpBuilder.setInsertionPointToStart(converter.getModuleOp().getBody()); |
3611 | auto mlirType = converter.genType(varType.declTypeSpec->derivedTypeSpec()); |
3612 | auto declMapperOp = firOpBuilder.create<mlir::omp::DeclareMapperOp>( |
3613 | loc, mapperNameStr, mlirType); |
3614 | auto ®ion = declMapperOp.getRegion(); |
3615 | firOpBuilder.createBlock(®ion); |
3616 | auto varVal = region.addArgument(firOpBuilder.getRefType(mlirType), loc); |
3617 | converter.bindSymbol(*varName.symbol, varVal); |
3618 | |
3619 | // Populate the declareMapper region with the map information. |
3620 | mlir::omp::DeclareMapperInfoOperands clauseOps; |
3621 | const auto *clauseList{ |
3622 | parser::Unwrap<parser::OmpClauseList>(declareMapperConstruct.t)}; |
3623 | List<Clause> clauses = makeClauses(*clauseList, semaCtx); |
3624 | ClauseProcessor cp(converter, semaCtx, clauses); |
3625 | cp.processMap(loc, stmtCtx, clauseOps); |
3626 | firOpBuilder.create<mlir::omp::DeclareMapperInfoOp>(loc, clauseOps.mapVars); |
3627 | } |
3628 | |
3629 | static void |
3630 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3631 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3632 | const parser::OpenMPDeclareTargetConstruct &declareTargetConstruct) { |
3633 | mlir::omp::DeclareTargetOperands clauseOps; |
3634 | llvm::SmallVector<DeclareTargetCapturePair> symbolAndClause; |
3635 | mlir::ModuleOp mod = converter.getFirOpBuilder().getModule(); |
3636 | getDeclareTargetInfo(converter, semaCtx, eval, declareTargetConstruct, |
3637 | clauseOps, symbolAndClause); |
3638 | |
3639 | for (const DeclareTargetCapturePair &symClause : symbolAndClause) { |
3640 | mlir::Operation *op = mod.lookupSymbol( |
3641 | converter.mangleName(std::get<const semantics::Symbol &>(symClause))); |
3642 | |
3643 | // Some symbols are deferred until later in the module, these are handled |
3644 | // upon finalization of the module for OpenMP inside of Bridge, so we simply |
3645 | // skip for now. |
3646 | if (!op) |
3647 | continue; |
3648 | |
3649 | markDeclareTarget( |
3650 | op, converter, |
3651 | std::get<mlir::omp::DeclareTargetCaptureClause>(symClause), |
3652 | clauseOps.deviceType); |
3653 | } |
3654 | } |
3655 | |
3656 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3657 | semantics::SemanticsContext &semaCtx, |
3658 | lower::pft::Evaluation &eval, |
3659 | const parser::OpenMPRequiresConstruct &requiresConstruct) { |
3660 | // Requires directives are gathered and processed in semantics and |
3661 | // then combined in the lowering bridge before triggering codegen |
3662 | // just once. Hence, there is no need to lower each individual |
3663 | // occurrence here. |
3664 | } |
3665 | |
3666 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3667 | semantics::SemanticsContext &semaCtx, |
3668 | lower::pft::Evaluation &eval, |
3669 | const parser::OpenMPThreadprivate &threadprivate) { |
3670 | // The directive is lowered when instantiating the variable to |
3671 | // support the case of threadprivate variable declared in module. |
3672 | } |
3673 | |
3674 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3675 | semantics::SemanticsContext &semaCtx, |
3676 | lower::pft::Evaluation &eval, |
3677 | const parser::OmpMetadirectiveDirective &meta) { |
3678 | TODO(converter.getCurrentLocation(), "METADIRECTIVE" ); |
3679 | } |
3680 | |
3681 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3682 | semantics::SemanticsContext &semaCtx, |
3683 | lower::pft::Evaluation &eval, |
3684 | const parser::OpenMPDeclarativeConstruct &ompDeclConstruct) { |
3685 | Fortran::common::visit( |
3686 | [&](auto &&s) { return genOMP(converter, symTable, semaCtx, eval, s); }, |
3687 | ompDeclConstruct.u); |
3688 | } |
3689 | |
3690 | //===----------------------------------------------------------------------===// |
3691 | // OpenMPStandaloneConstruct visitors |
3692 | //===----------------------------------------------------------------------===// |
3693 | |
3694 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3695 | semantics::SemanticsContext &semaCtx, |
3696 | lower::pft::Evaluation &eval, |
3697 | const parser::OpenMPSimpleStandaloneConstruct &construct) { |
3698 | const auto &directive = std::get<parser::OmpDirectiveName>(construct.v.t); |
3699 | List<Clause> clauses = makeClauses(construct.v.Clauses(), semaCtx); |
3700 | mlir::Location currentLocation = converter.genLocation(directive.source); |
3701 | |
3702 | ConstructQueue queue{ |
3703 | buildConstructQueue(converter.getFirOpBuilder().getModule(), semaCtx, |
3704 | eval, directive.source, directive.v, clauses)}; |
3705 | if (directive.v == llvm::omp::Directive::OMPD_ordered) { |
3706 | // Standalone "ordered" directive. |
3707 | genOrderedOp(converter, symTable, semaCtx, eval, currentLocation, queue, |
3708 | queue.begin()); |
3709 | } else { |
3710 | // Dispatch handles the "block-associated" variant of "ordered". |
3711 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
3712 | queue.begin()); |
3713 | } |
3714 | } |
3715 | |
3716 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3717 | semantics::SemanticsContext &semaCtx, |
3718 | lower::pft::Evaluation &eval, |
3719 | const parser::OpenMPFlushConstruct &construct) { |
3720 | const auto &argumentList = construct.v.Arguments(); |
3721 | const auto &clauseList = construct.v.Clauses(); |
3722 | ObjectList objects = makeObjects(argumentList, semaCtx); |
3723 | List<Clause> clauses = |
3724 | makeList(clauseList.v, [&](auto &&s) { return makeClause(s, semaCtx); }); |
3725 | mlir::Location currentLocation = converter.genLocation(construct.source); |
3726 | |
3727 | ConstructQueue queue{buildConstructQueue( |
3728 | converter.getFirOpBuilder().getModule(), semaCtx, eval, construct.source, |
3729 | llvm::omp::Directive::OMPD_flush, clauses)}; |
3730 | genFlushOp(converter, symTable, semaCtx, eval, currentLocation, objects, |
3731 | queue, queue.begin()); |
3732 | } |
3733 | |
3734 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3735 | semantics::SemanticsContext &semaCtx, |
3736 | lower::pft::Evaluation &eval, |
3737 | const parser::OpenMPCancelConstruct &cancelConstruct) { |
3738 | List<Clause> clauses = makeList(cancelConstruct.v.Clauses().v, [&](auto &&s) { |
3739 | return makeClause(s, semaCtx); |
3740 | }); |
3741 | mlir::Location loc = converter.genLocation(cancelConstruct.source); |
3742 | |
3743 | ConstructQueue queue{buildConstructQueue( |
3744 | converter.getFirOpBuilder().getModule(), semaCtx, eval, |
3745 | cancelConstruct.source, llvm::omp::Directive::OMPD_cancel, clauses)}; |
3746 | genCancelOp(converter, semaCtx, eval, loc, queue, queue.begin()); |
3747 | } |
3748 | |
3749 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3750 | semantics::SemanticsContext &semaCtx, |
3751 | lower::pft::Evaluation &eval, |
3752 | const parser::OpenMPCancellationPointConstruct |
3753 | &cancellationPointConstruct) { |
3754 | List<Clause> clauses = |
3755 | makeList(cancellationPointConstruct.v.Clauses().v, |
3756 | [&](auto &&s) { return makeClause(s, semaCtx); }); |
3757 | mlir::Location loc = converter.genLocation(cancellationPointConstruct.source); |
3758 | |
3759 | ConstructQueue queue{ |
3760 | buildConstructQueue(converter.getFirOpBuilder().getModule(), semaCtx, |
3761 | eval, cancellationPointConstruct.source, |
3762 | llvm::omp::Directive::OMPD_cancel, clauses)}; |
3763 | genCancellationPointOp(converter, semaCtx, eval, loc, queue, queue.begin()); |
3764 | } |
3765 | |
3766 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3767 | semantics::SemanticsContext &semaCtx, |
3768 | lower::pft::Evaluation &eval, |
3769 | const parser::OpenMPDepobjConstruct &construct) { |
3770 | // These values will be ignored until the construct itself is implemented, |
3771 | // but run them anyway for the sake of testing (via a Todo test). |
3772 | ObjectList objects = makeObjects(construct.v.Arguments(), semaCtx); |
3773 | assert(objects.size() == 1); |
3774 | List<Clause> clauses = makeClauses(construct.v.Clauses(), semaCtx); |
3775 | assert(clauses.size() == 1); |
3776 | (void)objects; |
3777 | (void)clauses; |
3778 | |
3779 | TODO(converter.getCurrentLocation(), "OpenMPDepobjConstruct" ); |
3780 | } |
3781 | |
3782 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3783 | semantics::SemanticsContext &semaCtx, |
3784 | lower::pft::Evaluation &eval, |
3785 | const parser::OpenMPInteropConstruct &interopConstruct) { |
3786 | TODO(converter.getCurrentLocation(), "OpenMPInteropConstruct" ); |
3787 | } |
3788 | |
3789 | static void |
3790 | genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3791 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
3792 | const parser::OpenMPStandaloneConstruct &standaloneConstruct) { |
3793 | Fortran::common::visit( |
3794 | [&](auto &&s) { return genOMP(converter, symTable, semaCtx, eval, s); }, |
3795 | standaloneConstruct.u); |
3796 | } |
3797 | |
3798 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3799 | semantics::SemanticsContext &semaCtx, |
3800 | lower::pft::Evaluation &eval, |
3801 | const parser::OpenMPAllocatorsConstruct &allocsConstruct) { |
3802 | TODO(converter.getCurrentLocation(), "OpenMPAllocatorsConstruct" ); |
3803 | } |
3804 | |
3805 | //===----------------------------------------------------------------------===// |
3806 | // OpenMPConstruct visitors |
3807 | //===----------------------------------------------------------------------===// |
3808 | |
3809 | [[maybe_unused]] static void |
3810 | dumpAtomicAnalysis(const parser::OpenMPAtomicConstruct::Analysis &analysis) { |
3811 | auto whatStr = [](int k) { |
3812 | std::string txt = "?" ; |
3813 | switch (k & parser::OpenMPAtomicConstruct::Analysis::Action) { |
3814 | case parser::OpenMPAtomicConstruct::Analysis::None: |
3815 | txt = "None" ; |
3816 | break; |
3817 | case parser::OpenMPAtomicConstruct::Analysis::Read: |
3818 | txt = "Read" ; |
3819 | break; |
3820 | case parser::OpenMPAtomicConstruct::Analysis::Write: |
3821 | txt = "Write" ; |
3822 | break; |
3823 | case parser::OpenMPAtomicConstruct::Analysis::Update: |
3824 | txt = "Update" ; |
3825 | break; |
3826 | } |
3827 | switch (k & parser::OpenMPAtomicConstruct::Analysis::Condition) { |
3828 | case parser::OpenMPAtomicConstruct::Analysis::IfTrue: |
3829 | txt += " | IfTrue" ; |
3830 | break; |
3831 | case parser::OpenMPAtomicConstruct::Analysis::IfFalse: |
3832 | txt += " | IfFalse" ; |
3833 | break; |
3834 | } |
3835 | return txt; |
3836 | }; |
3837 | |
3838 | auto exprStr = [&](const parser::TypedExpr &expr) { |
3839 | if (auto *maybe = expr.get()) { |
3840 | if (maybe->v) |
3841 | return maybe->v->AsFortran(); |
3842 | } |
3843 | return "<null>"s ; |
3844 | }; |
3845 | auto assignStr = [&](const parser::AssignmentStmt::TypedAssignment &assign) { |
3846 | if (auto *maybe = assign.get(); maybe && maybe->v) { |
3847 | std::string str; |
3848 | llvm::raw_string_ostream os(str); |
3849 | maybe->v->AsFortran(os); |
3850 | return str; |
3851 | } |
3852 | return "<null>"s ; |
3853 | }; |
3854 | |
3855 | const SomeExpr &atom = *analysis.atom.get()->v; |
3856 | |
3857 | llvm::errs() << "Analysis {\n" ; |
3858 | llvm::errs() << " atom: " << atom.AsFortran() << "\n" ; |
3859 | llvm::errs() << " cond: " << exprStr(analysis.cond) << "\n" ; |
3860 | llvm::errs() << " op0 {\n" ; |
3861 | llvm::errs() << " what: " << whatStr(analysis.op0.what) << "\n" ; |
3862 | llvm::errs() << " assign: " << assignStr(analysis.op0.assign) << "\n" ; |
3863 | llvm::errs() << " }\n" ; |
3864 | llvm::errs() << " op1 {\n" ; |
3865 | llvm::errs() << " what: " << whatStr(analysis.op1.what) << "\n" ; |
3866 | llvm::errs() << " assign: " << assignStr(analysis.op1.assign) << "\n" ; |
3867 | llvm::errs() << " }\n" ; |
3868 | llvm::errs() << "}\n" ; |
3869 | } |
3870 | |
3871 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3872 | semantics::SemanticsContext &semaCtx, |
3873 | lower::pft::Evaluation &eval, |
3874 | const parser::OpenMPAtomicConstruct &construct) { |
3875 | auto get = [](auto &&typedWrapper) -> decltype(&*typedWrapper.get()->v) { |
3876 | if (auto *maybe = typedWrapper.get(); maybe && maybe->v) { |
3877 | return &*maybe->v; |
3878 | } else { |
3879 | return nullptr; |
3880 | } |
3881 | }; |
3882 | |
3883 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
3884 | auto &dirSpec = std::get<parser::OmpDirectiveSpecification>(construct.t); |
3885 | List<Clause> clauses = makeClauses(dirSpec.Clauses(), semaCtx); |
3886 | lower::StatementContext stmtCtx; |
3887 | |
3888 | const parser::OpenMPAtomicConstruct::Analysis &analysis = construct.analysis; |
3889 | if (DumpAtomicAnalysis) |
3890 | dumpAtomicAnalysis(analysis); |
3891 | |
3892 | const semantics::SomeExpr &atom = *get(analysis.atom); |
3893 | mlir::Location loc = converter.genLocation(construct.source); |
3894 | mlir::Value atomAddr = |
3895 | fir::getBase(converter.genExprAddr(atom, stmtCtx, &loc)); |
3896 | mlir::IntegerAttr hint = getAtomicHint(converter, clauses); |
3897 | mlir::omp::ClauseMemoryOrderKindAttr memOrder = |
3898 | getAtomicMemoryOrder(converter, semaCtx, clauses); |
3899 | |
3900 | if (auto *cond = get(analysis.cond)) { |
3901 | (void)cond; |
3902 | TODO(loc, "OpenMP ATOMIC COMPARE" ); |
3903 | } else { |
3904 | int action0 = analysis.op0.what & analysis.Action; |
3905 | int action1 = analysis.op1.what & analysis.Action; |
3906 | mlir::Operation *captureOp = nullptr; |
3907 | fir::FirOpBuilder::InsertPoint preAt = builder.saveInsertionPoint(); |
3908 | fir::FirOpBuilder::InsertPoint atomicAt, postAt; |
3909 | |
3910 | if (construct.IsCapture()) { |
3911 | // Capturing operation. |
3912 | assert(action0 != analysis.None && action1 != analysis.None && |
3913 | "Expexcing two actions" ); |
3914 | (void)action0; |
3915 | (void)action1; |
3916 | captureOp = |
3917 | builder.create<mlir::omp::AtomicCaptureOp>(loc, hint, memOrder); |
3918 | // Set the non-atomic insertion point to before the atomic.capture. |
3919 | preAt = getInsertionPointBefore(captureOp); |
3920 | |
3921 | mlir::Block *block = builder.createBlock(&captureOp->getRegion(0)); |
3922 | builder.setInsertionPointToEnd(block); |
3923 | // Set the atomic insertion point to before the terminator inside |
3924 | // atomic.capture. |
3925 | mlir::Operation *term = builder.create<mlir::omp::TerminatorOp>(loc); |
3926 | atomicAt = getInsertionPointBefore(term); |
3927 | postAt = getInsertionPointAfter(captureOp); |
3928 | hint = nullptr; |
3929 | memOrder = nullptr; |
3930 | } else { |
3931 | // Non-capturing operation. |
3932 | assert(action0 != analysis.None && action1 == analysis.None && |
3933 | "Expexcing single action" ); |
3934 | assert(!(analysis.op0.what & analysis.Condition)); |
3935 | postAt = atomicAt = preAt; |
3936 | } |
3937 | |
3938 | // The builder's insertion point needs to be specifically set before |
3939 | // each call to `genAtomicOperation`. |
3940 | mlir::Operation *firstOp = genAtomicOperation( |
3941 | converter, loc, stmtCtx, analysis.op0.what, atomAddr, atom, |
3942 | *get(analysis.op0.assign), hint, memOrder, preAt, atomicAt, postAt); |
3943 | assert(firstOp && "Should have created an atomic operation" ); |
3944 | atomicAt = getInsertionPointAfter(firstOp); |
3945 | |
3946 | mlir::Operation *secondOp = nullptr; |
3947 | if (analysis.op1.what != analysis.None) { |
3948 | secondOp = genAtomicOperation(converter, loc, stmtCtx, analysis.op1.what, |
3949 | atomAddr, atom, *get(analysis.op1.assign), |
3950 | hint, memOrder, preAt, atomicAt, postAt); |
3951 | } |
3952 | |
3953 | if (construct.IsCapture()) { |
3954 | // If this is a capture operation, the first/second ops will be inside |
3955 | // of it. Set the insertion point to past the capture op itself. |
3956 | builder.restoreInsertionPoint(postAt); |
3957 | } else { |
3958 | if (secondOp) { |
3959 | builder.setInsertionPointAfter(secondOp); |
3960 | } else { |
3961 | builder.setInsertionPointAfter(firstOp); |
3962 | } |
3963 | } |
3964 | } |
3965 | } |
3966 | |
3967 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
3968 | semantics::SemanticsContext &semaCtx, |
3969 | lower::pft::Evaluation &eval, |
3970 | const parser::OpenMPBlockConstruct &blockConstruct) { |
3971 | const auto &beginBlockDirective = |
3972 | std::get<parser::OmpBeginBlockDirective>(blockConstruct.t); |
3973 | const auto &endBlockDirective = |
3974 | std::get<parser::OmpEndBlockDirective>(blockConstruct.t); |
3975 | mlir::Location currentLocation = |
3976 | converter.genLocation(beginBlockDirective.source); |
3977 | const auto origDirective = |
3978 | std::get<parser::OmpBlockDirective>(beginBlockDirective.t).v; |
3979 | List<Clause> clauses = makeClauses( |
3980 | std::get<parser::OmpClauseList>(beginBlockDirective.t), semaCtx); |
3981 | clauses.append(makeClauses( |
3982 | std::get<parser::OmpClauseList>(endBlockDirective.t), semaCtx)); |
3983 | |
3984 | assert(llvm::omp::blockConstructSet.test(origDirective) && |
3985 | "Expected block construct" ); |
3986 | (void)origDirective; |
3987 | |
3988 | for (const Clause &clause : clauses) { |
3989 | mlir::Location clauseLocation = converter.genLocation(clause.source); |
3990 | if (!std::holds_alternative<clause::Affinity>(clause.u) && |
3991 | !std::holds_alternative<clause::Allocate>(clause.u) && |
3992 | !std::holds_alternative<clause::Copyin>(clause.u) && |
3993 | !std::holds_alternative<clause::Copyprivate>(clause.u) && |
3994 | !std::holds_alternative<clause::Default>(clause.u) && |
3995 | !std::holds_alternative<clause::Defaultmap>(clause.u) && |
3996 | !std::holds_alternative<clause::Depend>(clause.u) && |
3997 | !std::holds_alternative<clause::Filter>(clause.u) && |
3998 | !std::holds_alternative<clause::Final>(clause.u) && |
3999 | !std::holds_alternative<clause::Firstprivate>(clause.u) && |
4000 | !std::holds_alternative<clause::HasDeviceAddr>(clause.u) && |
4001 | !std::holds_alternative<clause::If>(clause.u) && |
4002 | !std::holds_alternative<clause::IsDevicePtr>(clause.u) && |
4003 | !std::holds_alternative<clause::Map>(clause.u) && |
4004 | !std::holds_alternative<clause::Nowait>(clause.u) && |
4005 | !std::holds_alternative<clause::NumTeams>(clause.u) && |
4006 | !std::holds_alternative<clause::NumThreads>(clause.u) && |
4007 | !std::holds_alternative<clause::OmpxBare>(clause.u) && |
4008 | !std::holds_alternative<clause::Priority>(clause.u) && |
4009 | !std::holds_alternative<clause::Private>(clause.u) && |
4010 | !std::holds_alternative<clause::ProcBind>(clause.u) && |
4011 | !std::holds_alternative<clause::Reduction>(clause.u) && |
4012 | !std::holds_alternative<clause::Shared>(clause.u) && |
4013 | !std::holds_alternative<clause::Simd>(clause.u) && |
4014 | !std::holds_alternative<clause::ThreadLimit>(clause.u) && |
4015 | !std::holds_alternative<clause::Threads>(clause.u) && |
4016 | !std::holds_alternative<clause::UseDeviceAddr>(clause.u) && |
4017 | !std::holds_alternative<clause::UseDevicePtr>(clause.u) && |
4018 | !std::holds_alternative<clause::InReduction>(clause.u) && |
4019 | !std::holds_alternative<clause::Mergeable>(clause.u) && |
4020 | !std::holds_alternative<clause::Untied>(clause.u) && |
4021 | !std::holds_alternative<clause::TaskReduction>(clause.u) && |
4022 | !std::holds_alternative<clause::Detach>(clause.u)) { |
4023 | std::string name = |
4024 | parser::ToUpperCaseLetters(llvm::omp::getOpenMPClauseName(clause.id)); |
4025 | TODO(clauseLocation, name + " clause is not implemented yet" ); |
4026 | } |
4027 | } |
4028 | |
4029 | llvm::omp::Directive directive = |
4030 | std::get<parser::OmpBlockDirective>(beginBlockDirective.t).v; |
4031 | const parser::CharBlock &source = |
4032 | std::get<parser::OmpBlockDirective>(beginBlockDirective.t).source; |
4033 | ConstructQueue queue{ |
4034 | buildConstructQueue(converter.getFirOpBuilder().getModule(), semaCtx, |
4035 | eval, source, directive, clauses)}; |
4036 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
4037 | queue.begin()); |
4038 | } |
4039 | |
4040 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4041 | semantics::SemanticsContext &semaCtx, |
4042 | lower::pft::Evaluation &eval, |
4043 | const parser::OpenMPAssumeConstruct &assumeConstruct) { |
4044 | mlir::Location clauseLocation = converter.genLocation(assumeConstruct.source); |
4045 | TODO(clauseLocation, "OpenMP ASSUME construct" ); |
4046 | } |
4047 | |
4048 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4049 | semantics::SemanticsContext &semaCtx, |
4050 | lower::pft::Evaluation &eval, |
4051 | const parser::OpenMPCriticalConstruct &criticalConstruct) { |
4052 | const auto &cd = std::get<parser::OmpCriticalDirective>(criticalConstruct.t); |
4053 | List<Clause> clauses = |
4054 | makeClauses(std::get<parser::OmpClauseList>(cd.t), semaCtx); |
4055 | |
4056 | ConstructQueue queue{buildConstructQueue( |
4057 | converter.getFirOpBuilder().getModule(), semaCtx, eval, cd.source, |
4058 | llvm::omp::Directive::OMPD_critical, clauses)}; |
4059 | |
4060 | const auto &name = std::get<std::optional<parser::Name>>(cd.t); |
4061 | mlir::Location currentLocation = converter.getCurrentLocation(); |
4062 | genCriticalOp(converter, symTable, semaCtx, eval, currentLocation, queue, |
4063 | queue.begin(), name); |
4064 | } |
4065 | |
4066 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4067 | semantics::SemanticsContext &semaCtx, |
4068 | lower::pft::Evaluation &eval, |
4069 | const parser::OpenMPUtilityConstruct &) { |
4070 | TODO(converter.getCurrentLocation(), "OpenMPUtilityConstruct" ); |
4071 | } |
4072 | |
4073 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4074 | semantics::SemanticsContext &semaCtx, |
4075 | lower::pft::Evaluation &eval, |
4076 | const parser::OpenMPDispatchConstruct &) { |
4077 | TODO(converter.getCurrentLocation(), "OpenMPDispatchConstruct" ); |
4078 | } |
4079 | |
4080 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4081 | semantics::SemanticsContext &semaCtx, |
4082 | lower::pft::Evaluation &eval, |
4083 | const parser::OpenMPExecutableAllocate &execAllocConstruct) { |
4084 | TODO(converter.getCurrentLocation(), "OpenMPExecutableAllocate" ); |
4085 | } |
4086 | |
4087 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4088 | semantics::SemanticsContext &semaCtx, |
4089 | lower::pft::Evaluation &eval, |
4090 | const parser::OpenMPLoopConstruct &loopConstruct) { |
4091 | const auto &beginLoopDirective = |
4092 | std::get<parser::OmpBeginLoopDirective>(loopConstruct.t); |
4093 | List<Clause> clauses = makeClauses( |
4094 | std::get<parser::OmpClauseList>(beginLoopDirective.t), semaCtx); |
4095 | if (auto &endLoopDirective = |
4096 | std::get<std::optional<parser::OmpEndLoopDirective>>( |
4097 | loopConstruct.t)) { |
4098 | clauses.append(makeClauses( |
4099 | std::get<parser::OmpClauseList>(endLoopDirective->t), semaCtx)); |
4100 | } |
4101 | |
4102 | mlir::Location currentLocation = |
4103 | converter.genLocation(beginLoopDirective.source); |
4104 | |
4105 | llvm::omp::Directive directive = |
4106 | std::get<parser::OmpLoopDirective>(beginLoopDirective.t).v; |
4107 | const parser::CharBlock &source = |
4108 | std::get<parser::OmpLoopDirective>(beginLoopDirective.t).source; |
4109 | ConstructQueue queue{ |
4110 | buildConstructQueue(converter.getFirOpBuilder().getModule(), semaCtx, |
4111 | eval, source, directive, clauses)}; |
4112 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
4113 | queue.begin()); |
4114 | } |
4115 | |
4116 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4117 | semantics::SemanticsContext &semaCtx, |
4118 | lower::pft::Evaluation &eval, |
4119 | const parser::OpenMPSectionConstruct §ionConstruct) { |
4120 | // Do nothing here. SECTION is lowered inside of the lowering for Sections |
4121 | } |
4122 | |
4123 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4124 | semantics::SemanticsContext &semaCtx, |
4125 | lower::pft::Evaluation &eval, |
4126 | const parser::OpenMPSectionsConstruct §ionsConstruct) { |
4127 | const auto &beginSectionsDirective = |
4128 | std::get<parser::OmpBeginSectionsDirective>(sectionsConstruct.t); |
4129 | List<Clause> clauses = makeClauses( |
4130 | std::get<parser::OmpClauseList>(beginSectionsDirective.t), semaCtx); |
4131 | const auto &endSectionsDirective = |
4132 | std::get<parser::OmpEndSectionsDirective>(sectionsConstruct.t); |
4133 | const auto §ionBlocks = |
4134 | std::get<parser::OmpSectionBlocks>(sectionsConstruct.t); |
4135 | clauses.append(makeClauses( |
4136 | std::get<parser::OmpClauseList>(endSectionsDirective.t), semaCtx)); |
4137 | mlir::Location currentLocation = converter.getCurrentLocation(); |
4138 | |
4139 | llvm::omp::Directive directive = |
4140 | std::get<parser::OmpSectionsDirective>(beginSectionsDirective.t).v; |
4141 | const parser::CharBlock &source = |
4142 | std::get<parser::OmpSectionsDirective>(beginSectionsDirective.t).source; |
4143 | ConstructQueue queue{ |
4144 | buildConstructQueue(converter.getFirOpBuilder().getModule(), semaCtx, |
4145 | eval, source, directive, clauses)}; |
4146 | ConstructQueue::iterator next = queue.begin(); |
4147 | // Generate constructs that come first e.g. Parallel |
4148 | while (next != queue.end() && |
4149 | next->id != llvm::omp::Directive::OMPD_sections) { |
4150 | genOMPDispatch(converter, symTable, semaCtx, eval, currentLocation, queue, |
4151 | next); |
4152 | next = std::next(x: next); |
4153 | } |
4154 | |
4155 | // call genSectionsOp directly (not via genOMPDispatch) so that we can add the |
4156 | // sectionBlocks argument |
4157 | assert(next != queue.end()); |
4158 | assert(next->id == llvm::omp::Directive::OMPD_sections); |
4159 | genSectionsOp(converter, symTable, semaCtx, eval, currentLocation, queue, |
4160 | next, sectionBlocks); |
4161 | assert(std::next(next) == queue.end()); |
4162 | } |
4163 | |
4164 | static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable, |
4165 | semantics::SemanticsContext &semaCtx, |
4166 | lower::pft::Evaluation &eval, |
4167 | const parser::OpenMPConstruct &ompConstruct) { |
4168 | Fortran::common::visit( |
4169 | [&](auto &&s) { return genOMP(converter, symTable, semaCtx, eval, s); }, |
4170 | ompConstruct.u); |
4171 | } |
4172 | |
4173 | //===----------------------------------------------------------------------===// |
4174 | // Public functions |
4175 | //===----------------------------------------------------------------------===// |
4176 | |
4177 | mlir::Operation *Fortran::lower::genOpenMPTerminator(fir::FirOpBuilder &builder, |
4178 | mlir::Operation *op, |
4179 | mlir::Location loc) { |
4180 | if (mlir::isa<mlir::omp::AtomicUpdateOp, mlir::omp::DeclareReductionOp, |
4181 | mlir::omp::LoopNestOp>(op)) |
4182 | return builder.create<mlir::omp::YieldOp>(loc); |
4183 | return builder.create<mlir::omp::TerminatorOp>(loc); |
4184 | } |
4185 | |
4186 | void Fortran::lower::genOpenMPConstruct(lower::AbstractConverter &converter, |
4187 | lower::SymMap &symTable, |
4188 | semantics::SemanticsContext &semaCtx, |
4189 | lower::pft::Evaluation &eval, |
4190 | const parser::OpenMPConstruct &omp) { |
4191 | lower::SymMapScope scope(symTable); |
4192 | genOMP(converter, symTable, semaCtx, eval, omp); |
4193 | } |
4194 | |
4195 | void Fortran::lower::genOpenMPDeclarativeConstruct( |
4196 | lower::AbstractConverter &converter, lower::SymMap &symTable, |
4197 | semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval, |
4198 | const parser::OpenMPDeclarativeConstruct &omp) { |
4199 | genOMP(converter, symTable, semaCtx, eval, omp); |
4200 | genNestedEvaluations(converter, eval); |
4201 | } |
4202 | |
4203 | void Fortran::lower::genOpenMPSymbolProperties( |
4204 | lower::AbstractConverter &converter, const lower::pft::Variable &var) { |
4205 | assert(var.hasSymbol() && "Expecting Symbol" ); |
4206 | const semantics::Symbol &sym = var.getSymbol(); |
4207 | |
4208 | if (sym.test(semantics::Symbol::Flag::OmpThreadprivate)) |
4209 | lower::genThreadprivateOp(converter, var); |
4210 | |
4211 | if (sym.test(semantics::Symbol::Flag::OmpDeclareTarget)) |
4212 | lower::genDeclareTargetIntGlobal(converter, var); |
4213 | } |
4214 | |
4215 | int64_t |
4216 | Fortran::lower::getCollapseValue(const parser::OmpClauseList &clauseList) { |
4217 | for (const parser::OmpClause &clause : clauseList.v) { |
4218 | if (const auto &collapseClause = |
4219 | std::get_if<parser::OmpClause::Collapse>(&clause.u)) { |
4220 | const auto *expr = semantics::GetExpr(collapseClause->v); |
4221 | return evaluate::ToInt64(*expr).value(); |
4222 | } |
4223 | } |
4224 | return 1; |
4225 | } |
4226 | |
4227 | void Fortran::lower::genThreadprivateOp(lower::AbstractConverter &converter, |
4228 | const lower::pft::Variable &var) { |
4229 | fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder(); |
4230 | mlir::Location currentLocation = converter.getCurrentLocation(); |
4231 | |
4232 | const semantics::Symbol &sym = var.getSymbol(); |
4233 | mlir::Value symThreadprivateValue; |
4234 | if (const semantics::Symbol *common = |
4235 | semantics::FindCommonBlockContaining(sym.GetUltimate())) { |
4236 | mlir::Value commonValue = converter.getSymbolAddress(*common); |
4237 | if (mlir::isa<mlir::omp::ThreadprivateOp>(commonValue.getDefiningOp())) { |
4238 | // Generate ThreadprivateOp for a common block instead of its members and |
4239 | // only do it once for a common block. |
4240 | return; |
4241 | } |
4242 | // Generate ThreadprivateOp and rebind the common block. |
4243 | mlir::Value commonThreadprivateValue = |
4244 | firOpBuilder.create<mlir::omp::ThreadprivateOp>( |
4245 | currentLocation, commonValue.getType(), commonValue); |
4246 | converter.bindSymbol(*common, commonThreadprivateValue); |
4247 | // Generate the threadprivate value for the common block member. |
4248 | symThreadprivateValue = genCommonBlockMember(converter, currentLocation, |
4249 | sym, commonThreadprivateValue); |
4250 | } else if (!var.isGlobal()) { |
4251 | // Non-global variable which can be in threadprivate directive must be one |
4252 | // variable in main program, and it has implicit SAVE attribute. Take it as |
4253 | // with SAVE attribute, so to create GlobalOp for it to simplify the |
4254 | // translation to LLVM IR. |
4255 | // Avoids performing multiple globalInitializations. |
4256 | fir::GlobalOp global; |
4257 | auto module = converter.getModuleOp(); |
4258 | std::string globalName = converter.mangleName(sym); |
4259 | if (module.lookupSymbol<fir::GlobalOp>(globalName)) |
4260 | global = module.lookupSymbol<fir::GlobalOp>(globalName); |
4261 | else |
4262 | global = globalInitialization(converter, firOpBuilder, sym, var, |
4263 | currentLocation); |
4264 | |
4265 | mlir::Value symValue = firOpBuilder.create<fir::AddrOfOp>( |
4266 | currentLocation, global.resultType(), global.getSymbol()); |
4267 | symThreadprivateValue = firOpBuilder.create<mlir::omp::ThreadprivateOp>( |
4268 | currentLocation, symValue.getType(), symValue); |
4269 | } else { |
4270 | mlir::Value symValue = converter.getSymbolAddress(sym); |
4271 | |
4272 | // The symbol may be use-associated multiple times, and nothing needs to be |
4273 | // done after the original symbol is mapped to the threadprivatized value |
4274 | // for the first time. Use the threadprivatized value directly. |
4275 | mlir::Operation *op; |
4276 | if (auto declOp = symValue.getDefiningOp<hlfir::DeclareOp>()) |
4277 | op = declOp.getMemref().getDefiningOp(); |
4278 | else |
4279 | op = symValue.getDefiningOp(); |
4280 | if (mlir::isa<mlir::omp::ThreadprivateOp>(op)) |
4281 | return; |
4282 | |
4283 | symThreadprivateValue = firOpBuilder.create<mlir::omp::ThreadprivateOp>( |
4284 | currentLocation, symValue.getType(), symValue); |
4285 | } |
4286 | |
4287 | fir::ExtendedValue sexv = converter.getSymbolExtendedValue(sym); |
4288 | fir::ExtendedValue symThreadprivateExv = |
4289 | getExtendedValue(sexv, symThreadprivateValue); |
4290 | converter.bindSymbol(sym, symThreadprivateExv); |
4291 | } |
4292 | |
4293 | // This function replicates threadprivate's behaviour of generating |
4294 | // an internal fir.GlobalOp for non-global variables in the main program |
4295 | // that have the implicit SAVE attribute, to simplifiy LLVM-IR and MLIR |
4296 | // generation. |
4297 | void Fortran::lower::genDeclareTargetIntGlobal( |
4298 | lower::AbstractConverter &converter, const lower::pft::Variable &var) { |
4299 | if (!var.isGlobal()) { |
4300 | // A non-global variable which can be in a declare target directive must |
4301 | // be a variable in the main program, and it has the implicit SAVE |
4302 | // attribute. We create a GlobalOp for it to simplify the translation to |
4303 | // LLVM IR. |
4304 | globalInitialization(converter, converter.getFirOpBuilder(), |
4305 | var.getSymbol(), var, converter.getCurrentLocation()); |
4306 | } |
4307 | } |
4308 | |
4309 | bool Fortran::lower::isOpenMPTargetConstruct( |
4310 | const parser::OpenMPConstruct &omp) { |
4311 | llvm::omp::Directive dir = llvm::omp::Directive::OMPD_unknown; |
4312 | if (const auto *block = std::get_if<parser::OpenMPBlockConstruct>(&omp.u)) { |
4313 | const auto &begin = std::get<parser::OmpBeginBlockDirective>(block->t); |
4314 | dir = std::get<parser::OmpBlockDirective>(begin.t).v; |
4315 | } else if (const auto *loop = |
4316 | std::get_if<parser::OpenMPLoopConstruct>(&omp.u)) { |
4317 | const auto &begin = std::get<parser::OmpBeginLoopDirective>(loop->t); |
4318 | dir = std::get<parser::OmpLoopDirective>(begin.t).v; |
4319 | } |
4320 | return llvm::omp::allTargetSet.test(dir); |
4321 | } |
4322 | |
4323 | void Fortran::lower::gatherOpenMPDeferredDeclareTargets( |
4324 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
4325 | lower::pft::Evaluation &eval, |
4326 | const parser::OpenMPDeclarativeConstruct &ompDecl, |
4327 | llvm::SmallVectorImpl<OMPDeferredDeclareTargetInfo> |
4328 | &deferredDeclareTarget) { |
4329 | Fortran::common::visit( |
4330 | common::visitors{ |
4331 | [&](const parser::OpenMPDeclareTargetConstruct &ompReq) { |
4332 | collectDeferredDeclareTargets(converter, semaCtx, eval, ompReq, |
4333 | deferredDeclareTarget); |
4334 | }, |
4335 | [&](const auto &) {}, |
4336 | }, |
4337 | ompDecl.u); |
4338 | } |
4339 | |
4340 | bool Fortran::lower::isOpenMPDeviceDeclareTarget( |
4341 | lower::AbstractConverter &converter, semantics::SemanticsContext &semaCtx, |
4342 | lower::pft::Evaluation &eval, |
4343 | const parser::OpenMPDeclarativeConstruct &ompDecl) { |
4344 | return Fortran::common::visit( |
4345 | common::visitors{ |
4346 | [&](const parser::OpenMPDeclareTargetConstruct &ompReq) { |
4347 | mlir::omp::DeclareTargetDeviceType targetType = |
4348 | getDeclareTargetFunctionDevice(converter, semaCtx, eval, ompReq) |
4349 | .value_or(mlir::omp::DeclareTargetDeviceType::host); |
4350 | return targetType != mlir::omp::DeclareTargetDeviceType::host; |
4351 | }, |
4352 | [&](const auto &) { return false; }, |
4353 | }, |
4354 | ompDecl.u); |
4355 | } |
4356 | |
4357 | // In certain cases such as subroutine or function interfaces which declare |
4358 | // but do not define or directly call the subroutine or function in the same |
4359 | // module, their lowering is delayed until after the declare target construct |
4360 | // itself is processed, so there symbol is not within the table. |
4361 | // |
4362 | // This function will also return true if we encounter any device declare |
4363 | // target cases, to satisfy checking if we require the requires attributes |
4364 | // on the module. |
4365 | bool Fortran::lower::markOpenMPDeferredDeclareTargetFunctions( |
4366 | mlir::Operation *mod, |
4367 | llvm::SmallVectorImpl<OMPDeferredDeclareTargetInfo> &deferredDeclareTargets, |
4368 | AbstractConverter &converter) { |
4369 | bool deviceCodeFound = false; |
4370 | auto modOp = llvm::cast<mlir::ModuleOp>(mod); |
4371 | for (auto declTar : deferredDeclareTargets) { |
4372 | mlir::Operation *op = modOp.lookupSymbol(converter.mangleName(declTar.sym)); |
4373 | |
4374 | // Due to interfaces being optionally emitted on usage in a module, |
4375 | // not finding an operation at this point cannot be a hard error, we |
4376 | // simply ignore it for now. |
4377 | // TODO: Add semantic checks for detecting cases where an erronous |
4378 | // (undefined) symbol has been supplied to a declare target clause |
4379 | if (!op) |
4380 | continue; |
4381 | |
4382 | auto devType = declTar.declareTargetDeviceType; |
4383 | if (!deviceCodeFound && devType != mlir::omp::DeclareTargetDeviceType::host) |
4384 | deviceCodeFound = true; |
4385 | |
4386 | markDeclareTarget(op, converter, declTar.declareTargetCaptureClause, |
4387 | devType); |
4388 | } |
4389 | |
4390 | return deviceCodeFound; |
4391 | } |
4392 | |
4393 | void Fortran::lower::genOpenMPRequires(mlir::Operation *mod, |
4394 | const semantics::Symbol *symbol) { |
4395 | using MlirRequires = mlir::omp::ClauseRequires; |
4396 | using SemaRequires = semantics::WithOmpDeclarative::RequiresFlag; |
4397 | |
4398 | if (auto offloadMod = |
4399 | llvm::dyn_cast<mlir::omp::OffloadModuleInterface>(mod)) { |
4400 | semantics::WithOmpDeclarative::RequiresFlags semaFlags; |
4401 | if (symbol) { |
4402 | common::visit( |
4403 | [&](const auto &details) { |
4404 | if constexpr (std::is_base_of_v<semantics::WithOmpDeclarative, |
4405 | std::decay_t<decltype(details)>>) { |
4406 | if (details.has_ompRequires()) |
4407 | semaFlags = *details.ompRequires(); |
4408 | } |
4409 | }, |
4410 | symbol->details()); |
4411 | } |
4412 | |
4413 | // Use pre-populated omp.requires module attribute if it was set, so that |
4414 | // the "-fopenmp-force-usm" compiler option is honored. |
4415 | MlirRequires mlirFlags = offloadMod.getRequires(); |
4416 | if (semaFlags.test(SemaRequires::ReverseOffload)) |
4417 | mlirFlags = mlirFlags | MlirRequires::reverse_offload; |
4418 | if (semaFlags.test(SemaRequires::UnifiedAddress)) |
4419 | mlirFlags = mlirFlags | MlirRequires::unified_address; |
4420 | if (semaFlags.test(SemaRequires::UnifiedSharedMemory)) |
4421 | mlirFlags = mlirFlags | MlirRequires::unified_shared_memory; |
4422 | if (semaFlags.test(SemaRequires::DynamicAllocators)) |
4423 | mlirFlags = mlirFlags | MlirRequires::dynamic_allocators; |
4424 | |
4425 | offloadMod.setRequires(mlirFlags); |
4426 | } |
4427 | } |
4428 | |