1 | //===-- ConvertCall.cpp ---------------------------------------------------===// |
---|---|
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/ |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "flang/Lower/ConvertCall.h" |
14 | #include "flang/Lower/Allocatable.h" |
15 | #include "flang/Lower/ConvertExprToHLFIR.h" |
16 | #include "flang/Lower/ConvertProcedureDesignator.h" |
17 | #include "flang/Lower/ConvertVariable.h" |
18 | #include "flang/Lower/CustomIntrinsicCall.h" |
19 | #include "flang/Lower/HlfirIntrinsics.h" |
20 | #include "flang/Lower/StatementContext.h" |
21 | #include "flang/Lower/SymbolMap.h" |
22 | #include "flang/Optimizer/Builder/BoxValue.h" |
23 | #include "flang/Optimizer/Builder/Character.h" |
24 | #include "flang/Optimizer/Builder/FIRBuilder.h" |
25 | #include "flang/Optimizer/Builder/HLFIRTools.h" |
26 | #include "flang/Optimizer/Builder/IntrinsicCall.h" |
27 | #include "flang/Optimizer/Builder/LowLevelIntrinsics.h" |
28 | #include "flang/Optimizer/Builder/MutableBox.h" |
29 | #include "flang/Optimizer/Builder/Runtime/Derived.h" |
30 | #include "flang/Optimizer/Builder/Todo.h" |
31 | #include "flang/Optimizer/Dialect/CUF/CUFOps.h" |
32 | #include "flang/Optimizer/Dialect/FIROpsSupport.h" |
33 | #include "flang/Optimizer/HLFIR/HLFIROps.h" |
34 | #include "mlir/IR/IRMapping.h" |
35 | #include "llvm/ADT/TypeSwitch.h" |
36 | #include "llvm/Support/CommandLine.h" |
37 | #include "llvm/Support/Debug.h" |
38 | #include <optional> |
39 | |
40 | #define DEBUG_TYPE "flang-lower-expr" |
41 | |
42 | static llvm::cl::opt<bool> useHlfirIntrinsicOps( |
43 | "use-hlfir-intrinsic-ops", llvm::cl::init(Val: true), |
44 | llvm::cl::desc("Lower via HLFIR transformational intrinsic operations such " |
45 | "as hlfir.sum")); |
46 | |
47 | static constexpr char tempResultName[] = ".tmp.func_result"; |
48 | |
49 | /// Helper to package a Value and its properties into an ExtendedValue. |
50 | static fir::ExtendedValue toExtendedValue(mlir::Location loc, mlir::Value base, |
51 | llvm::ArrayRef<mlir::Value> extents, |
52 | llvm::ArrayRef<mlir::Value> lengths) { |
53 | mlir::Type type = base.getType(); |
54 | if (mlir::isa<fir::BaseBoxType>(type)) |
55 | return fir::BoxValue(base, /*lbounds=*/{}, lengths, extents); |
56 | type = fir::unwrapRefType(type); |
57 | if (mlir::isa<fir::BaseBoxType>(type)) |
58 | return fir::MutableBoxValue(base, lengths, /*mutableProperties*/ {}); |
59 | if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(type)) { |
60 | if (seqTy.getDimension() != extents.size()) |
61 | fir::emitFatalError(loc, "incorrect number of extents for array"); |
62 | if (mlir::isa<fir::CharacterType>(seqTy.getEleTy())) { |
63 | if (lengths.empty()) |
64 | fir::emitFatalError(loc, "missing length for character"); |
65 | assert(lengths.size() == 1); |
66 | return fir::CharArrayBoxValue(base, lengths[0], extents); |
67 | } |
68 | return fir::ArrayBoxValue(base, extents); |
69 | } |
70 | if (mlir::isa<fir::CharacterType>(type)) { |
71 | if (lengths.empty()) |
72 | fir::emitFatalError(loc, "missing length for character"); |
73 | assert(lengths.size() == 1); |
74 | return fir::CharBoxValue(base, lengths[0]); |
75 | } |
76 | return base; |
77 | } |
78 | |
79 | /// Lower a type(C_PTR/C_FUNPTR) argument with VALUE attribute into a |
80 | /// reference. A C pointer can correspond to a Fortran dummy argument of type |
81 | /// C_PTR with the VALUE attribute. (see 18.3.6 note 3). |
82 | static mlir::Value genRecordCPtrValueArg(fir::FirOpBuilder &builder, |
83 | mlir::Location loc, mlir::Value rec, |
84 | mlir::Type ty) { |
85 | mlir::Value cAddr = fir::factory::genCPtrOrCFunptrAddr(builder, loc, rec, ty); |
86 | mlir::Value cVal = builder.create<fir::LoadOp>(loc, cAddr); |
87 | return builder.createConvert(loc, cAddr.getType(), cVal); |
88 | } |
89 | |
90 | // Find the argument that corresponds to the host associations. |
91 | // Verify some assumptions about how the signature was built here. |
92 | [[maybe_unused]] static unsigned findHostAssocTuplePos(mlir::func::FuncOp fn) { |
93 | // Scan the argument list from last to first as the host associations are |
94 | // appended for now. |
95 | for (unsigned i = fn.getNumArguments(); i > 0; --i) |
96 | if (fn.getArgAttr(i - 1, fir::getHostAssocAttrName())) { |
97 | // Host assoc tuple must be last argument (for now). |
98 | assert(i == fn.getNumArguments() && "tuple must be last"); |
99 | return i - 1; |
100 | } |
101 | llvm_unreachable("anyFuncArgsHaveAttr failed"); |
102 | } |
103 | |
104 | mlir::Value |
105 | Fortran::lower::argumentHostAssocs(Fortran::lower::AbstractConverter &converter, |
106 | mlir::Value arg) { |
107 | if (auto addr = mlir::dyn_cast_or_null<fir::AddrOfOp>(arg.getDefiningOp())) { |
108 | auto &builder = converter.getFirOpBuilder(); |
109 | if (auto funcOp = builder.getNamedFunction(addr.getSymbol())) |
110 | if (fir::anyFuncArgsHaveAttr(funcOp, fir::getHostAssocAttrName())) |
111 | return converter.hostAssocTupleValue(); |
112 | } |
113 | return {}; |
114 | } |
115 | |
116 | static bool mustCastFuncOpToCopeWithImplicitInterfaceMismatch( |
117 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
118 | mlir::FunctionType callSiteType, mlir::FunctionType funcOpType) { |
119 | // Deal with argument number mismatch by making a function pointer so |
120 | // that function type cast can be inserted. Do not emit a warning here |
121 | // because this can happen in legal program if the function is not |
122 | // defined here and it was first passed as an argument without any more |
123 | // information. |
124 | if (callSiteType.getNumResults() != funcOpType.getNumResults() || |
125 | callSiteType.getNumInputs() != funcOpType.getNumInputs()) |
126 | return true; |
127 | |
128 | // Implicit interface result type mismatch are not standard Fortran, but |
129 | // some compilers are not complaining about it. The front end is not |
130 | // protecting lowering from this currently. Support this with a |
131 | // discouraging warning. |
132 | // Cast the actual function to the current caller implicit type because |
133 | // that is the behavior we would get if we could not see the definition. |
134 | if (callSiteType.getResults() != funcOpType.getResults()) { |
135 | LLVM_DEBUG(mlir::emitWarning( |
136 | loc, "a return type mismatch is not standard compliant and may " |
137 | "lead to undefined behavior.")); |
138 | return true; |
139 | } |
140 | |
141 | // In HLFIR, there is little attempt to cope with implicit interface |
142 | // mismatch on the arguments. The argument are always prepared according |
143 | // to the implicit interface. Cast the actual function if any of the |
144 | // argument mismatch cannot be dealt with a simple fir.convert. |
145 | if (converter.getLoweringOptions().getLowerToHighLevelFIR()) |
146 | for (auto [actualType, dummyType] : |
147 | llvm::zip(callSiteType.getInputs(), funcOpType.getInputs())) |
148 | if (actualType != dummyType && |
149 | !fir::ConvertOp::canBeConverted(actualType, dummyType)) |
150 | return true; |
151 | return false; |
152 | } |
153 | |
154 | static mlir::Value readDim3Value(fir::FirOpBuilder &builder, mlir::Location loc, |
155 | mlir::Value dim3Addr, llvm::StringRef comp) { |
156 | mlir::Type i32Ty = builder.getI32Type(); |
157 | mlir::Type refI32Ty = fir::ReferenceType::get(i32Ty); |
158 | llvm::SmallVector<mlir::Value> lenParams; |
159 | |
160 | mlir::Value designate = builder.create<hlfir::DesignateOp>( |
161 | loc, refI32Ty, dim3Addr, /*component=*/comp, |
162 | /*componentShape=*/mlir::Value{}, hlfir::DesignateOp::Subscripts{}, |
163 | /*substring=*/mlir::ValueRange{}, /*complexPartAttr=*/std::nullopt, |
164 | mlir::Value{}, lenParams); |
165 | |
166 | return hlfir::loadTrivialScalar(loc, builder, hlfir::Entity{designate}); |
167 | } |
168 | |
169 | static mlir::Value remapActualToDummyDescriptor( |
170 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
171 | Fortran::lower::SymMap &symMap, |
172 | const Fortran::lower::CallerInterface::PassedEntity &arg, |
173 | Fortran::lower::CallerInterface &caller, bool isBindcCall) { |
174 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
175 | mlir::IndexType idxTy = builder.getIndexType(); |
176 | mlir::Value zero = builder.createIntegerConstant(loc, idxTy, 0); |
177 | Fortran::lower::StatementContext localStmtCtx; |
178 | auto lowerSpecExpr = [&](const auto &expr, |
179 | bool isAssumedSizeExtent) -> mlir::Value { |
180 | mlir::Value convertExpr = builder.createConvert( |
181 | loc, idxTy, fir::getBase(converter.genExprValue(expr, localStmtCtx))); |
182 | if (isAssumedSizeExtent) |
183 | return convertExpr; |
184 | return fir::factory::genMaxWithZero(builder, loc, convertExpr); |
185 | }; |
186 | bool mapSymbols = caller.mustMapInterfaceSymbolsForDummyArgument(arg); |
187 | if (mapSymbols) { |
188 | symMap.pushScope(); |
189 | const Fortran::semantics::Symbol *sym = caller.getDummySymbol(arg); |
190 | assert(sym && "call must have explicit interface to map interface symbols"); |
191 | Fortran::lower::mapCallInterfaceSymbolsForDummyArgument(converter, caller, |
192 | symMap, *sym); |
193 | } |
194 | llvm::SmallVector<mlir::Value> extents; |
195 | llvm::SmallVector<mlir::Value> lengths; |
196 | mlir::Type dummyBoxType = caller.getDummyArgumentType(arg); |
197 | mlir::Type dummyBaseType = fir::unwrapPassByRefType(dummyBoxType); |
198 | if (mlir::isa<fir::SequenceType>(dummyBaseType)) |
199 | caller.walkDummyArgumentExtents( |
200 | arg, [&](const Fortran::lower::SomeExpr &e, bool isAssumedSizeExtent) { |
201 | extents.emplace_back(lowerSpecExpr(e, isAssumedSizeExtent)); |
202 | }); |
203 | mlir::Value shape; |
204 | if (!extents.empty()) { |
205 | if (isBindcCall) { |
206 | // Preserve zero lower bounds (see F'2023 18.5.3). |
207 | llvm::SmallVector<mlir::Value> lowerBounds(extents.size(), zero); |
208 | shape = builder.genShape(loc, lowerBounds, extents); |
209 | } else { |
210 | shape = builder.genShape(loc, extents); |
211 | } |
212 | } |
213 | |
214 | hlfir::Entity explicitArgument = hlfir::Entity{caller.getInput(arg)}; |
215 | mlir::Type dummyElementType = fir::unwrapSequenceType(dummyBaseType); |
216 | if (auto recType = llvm::dyn_cast<fir::RecordType>(dummyElementType)) |
217 | if (recType.getNumLenParams() > 0) |
218 | TODO(loc, "sequence association of length parameterized derived type " |
219 | "dummy arguments"); |
220 | if (fir::isa_char(dummyElementType)) |
221 | lengths.emplace_back(hlfir::genCharLength(loc, builder, explicitArgument)); |
222 | mlir::Value baseAddr = |
223 | hlfir::genVariableRawAddress(loc, builder, explicitArgument); |
224 | baseAddr = builder.createConvert(loc, fir::ReferenceType::get(dummyBaseType), |
225 | baseAddr); |
226 | mlir::Value mold; |
227 | if (fir::isPolymorphicType(dummyBoxType)) |
228 | mold = explicitArgument; |
229 | mlir::Value remapped = |
230 | builder.create<fir::EmboxOp>(loc, dummyBoxType, baseAddr, shape, |
231 | /*slice=*/mlir::Value{}, lengths, mold); |
232 | if (mapSymbols) |
233 | symMap.popScope(); |
234 | return remapped; |
235 | } |
236 | |
237 | /// Create a descriptor for sequenced associated descriptor that are passed |
238 | /// by descriptor. Sequence association (F'2023 15.5.2.12) implies that the |
239 | /// dummy shape and rank need to not be the same as the actual argument. This |
240 | /// helper creates a descriptor based on the dummy shape and rank (sequence |
241 | /// association can only happen with explicit and assumed-size array) so that it |
242 | /// is safe to assume the rank of the incoming descriptor inside the callee. |
243 | /// This helper must be called once all the actual arguments have been lowered |
244 | /// and placed inside "caller". Copy-in/copy-out must already have been |
245 | /// generated if needed using the actual argument shape (the dummy shape may be |
246 | /// assumed-size). |
247 | static void remapActualToDummyDescriptors( |
248 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
249 | Fortran::lower::SymMap &symMap, |
250 | const Fortran::lower::PreparedActualArguments &loweredActuals, |
251 | Fortran::lower::CallerInterface &caller, bool isBindcCall) { |
252 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
253 | for (auto [preparedActual, arg] : |
254 | llvm::zip(loweredActuals, caller.getPassedArguments())) { |
255 | if (arg.isSequenceAssociatedDescriptor()) { |
256 | if (!preparedActual.value().handleDynamicOptional()) { |
257 | mlir::Value remapped = remapActualToDummyDescriptor( |
258 | loc, converter, symMap, arg, caller, isBindcCall); |
259 | caller.placeInput(arg, remapped); |
260 | } else { |
261 | // Absent optional actual argument descriptor cannot be read and |
262 | // remapped unconditionally. |
263 | mlir::Type dummyType = caller.getDummyArgumentType(arg); |
264 | mlir::Value isPresent = preparedActual.value().getIsPresent(); |
265 | auto &argLambdaCapture = arg; |
266 | mlir::Value remapped = |
267 | builder |
268 | .genIfOp(loc, {dummyType}, isPresent, |
269 | /*withElseRegion=*/true) |
270 | .genThen([&]() { |
271 | mlir::Value newBox = remapActualToDummyDescriptor( |
272 | loc, converter, symMap, argLambdaCapture, caller, |
273 | isBindcCall); |
274 | builder.create<fir::ResultOp>(loc, newBox); |
275 | }) |
276 | .genElse([&]() { |
277 | mlir::Value absent = |
278 | builder.create<fir::AbsentOp>(loc, dummyType); |
279 | builder.create<fir::ResultOp>(loc, absent); |
280 | }) |
281 | .getResults()[0]; |
282 | caller.placeInput(arg, remapped); |
283 | } |
284 | } |
285 | } |
286 | } |
287 | |
288 | std::pair<Fortran::lower::LoweredResult, bool> |
289 | Fortran::lower::genCallOpAndResult( |
290 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
291 | Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx, |
292 | Fortran::lower::CallerInterface &caller, mlir::FunctionType callSiteType, |
293 | std::optional<mlir::Type> resultType, bool isElemental) { |
294 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
295 | using PassBy = Fortran::lower::CallerInterface::PassEntityBy; |
296 | bool mustPopSymMap = false; |
297 | if (caller.mustMapInterfaceSymbolsForResult()) { |
298 | symMap.pushScope(); |
299 | mustPopSymMap = true; |
300 | Fortran::lower::mapCallInterfaceSymbolsForResult(converter, caller, symMap); |
301 | } |
302 | // If this is an indirect call, retrieve the function address. Also retrieve |
303 | // the result length if this is a character function (note that this length |
304 | // will be used only if there is no explicit length in the local interface). |
305 | mlir::Value funcPointer; |
306 | mlir::Value charFuncPointerLength; |
307 | if (const Fortran::evaluate::ProcedureDesignator *procDesignator = |
308 | caller.getIfIndirectCall()) { |
309 | if (mlir::Value passedArg = caller.getIfPassedArg()) { |
310 | // Procedure pointer component call with PASS argument. To avoid |
311 | // "double" lowering of the ComponentRef, semantics only place the |
312 | // ComponentRef in the ActualArguments, not in the ProcedureDesignator ( |
313 | // that is only the component symbol). |
314 | // Fetch the passed argument and addresses of its procedure pointer |
315 | // component. |
316 | funcPointer = Fortran::lower::derefPassProcPointerComponent( |
317 | loc, converter, *procDesignator, passedArg, symMap, stmtCtx); |
318 | } else { |
319 | Fortran::lower::SomeExpr expr{*procDesignator}; |
320 | fir::ExtendedValue loweredProc = |
321 | converter.genExprAddr(loc, expr, stmtCtx); |
322 | funcPointer = fir::getBase(loweredProc); |
323 | // Dummy procedure may have assumed length, in which case the result |
324 | // length was passed along the dummy procedure. |
325 | // This is not possible with procedure pointer components. |
326 | if (const fir::CharBoxValue *charBox = loweredProc.getCharBox()) |
327 | charFuncPointerLength = charBox->getLen(); |
328 | } |
329 | } |
330 | const bool isExprCall = |
331 | converter.getLoweringOptions().getLowerToHighLevelFIR() && |
332 | callSiteType.getNumResults() == 1 && |
333 | llvm::isa<fir::SequenceType>(callSiteType.getResult(0)); |
334 | |
335 | mlir::IndexType idxTy = builder.getIndexType(); |
336 | auto lowerSpecExpr = [&](const auto &expr) -> mlir::Value { |
337 | mlir::Value convertExpr = builder.createConvert( |
338 | loc, idxTy, fir::getBase(converter.genExprValue(expr, stmtCtx))); |
339 | return fir::factory::genMaxWithZero(builder, loc, convertExpr); |
340 | }; |
341 | llvm::SmallVector<mlir::Value> resultLengths; |
342 | mlir::Value arrayResultShape; |
343 | hlfir::EvaluateInMemoryOp evaluateInMemory; |
344 | auto allocatedResult = [&]() -> std::optional<fir::ExtendedValue> { |
345 | llvm::SmallVector<mlir::Value> extents; |
346 | llvm::SmallVector<mlir::Value> lengths; |
347 | if (!caller.callerAllocateResult()) |
348 | return {}; |
349 | mlir::Type type = caller.getResultStorageType(); |
350 | if (mlir::isa<fir::SequenceType>(type)) |
351 | caller.walkResultExtents( |
352 | [&](const Fortran::lower::SomeExpr &e, bool isAssumedSizeExtent) { |
353 | assert(!isAssumedSizeExtent && "result cannot be assumed-size"); |
354 | extents.emplace_back(lowerSpecExpr(e)); |
355 | }); |
356 | caller.walkResultLengths( |
357 | [&](const Fortran::lower::SomeExpr &e, bool isAssumedSizeExtent) { |
358 | assert(!isAssumedSizeExtent && "result cannot be assumed-size"); |
359 | lengths.emplace_back(lowerSpecExpr(e)); |
360 | }); |
361 | |
362 | // Result length parameters should not be provided to box storage |
363 | // allocation and save_results, but they are still useful information to |
364 | // keep in the ExtendedValue if non-deferred. |
365 | if (!mlir::isa<fir::BoxType>(type)) { |
366 | if (fir::isa_char(fir::unwrapSequenceType(type)) && lengths.empty()) { |
367 | // Calling an assumed length function. This is only possible if this |
368 | // is a call to a character dummy procedure. |
369 | if (!charFuncPointerLength) |
370 | fir::emitFatalError(loc, "failed to retrieve character function " |
371 | "length while calling it"); |
372 | lengths.push_back(charFuncPointerLength); |
373 | } |
374 | resultLengths = lengths; |
375 | } |
376 | |
377 | if (!extents.empty()) |
378 | arrayResultShape = builder.genShape(loc, extents); |
379 | |
380 | if (isExprCall) { |
381 | mlir::Type exprType = hlfir::getExprType(type); |
382 | evaluateInMemory = builder.create<hlfir::EvaluateInMemoryOp>( |
383 | loc, exprType, arrayResultShape, resultLengths); |
384 | builder.setInsertionPointToStart(&evaluateInMemory.getBody().front()); |
385 | return toExtendedValue(loc, evaluateInMemory.getMemory(), extents, |
386 | lengths); |
387 | } |
388 | |
389 | if ((!extents.empty() || !lengths.empty()) && !isElemental) { |
390 | // Note: in the elemental context, the alloca ownership inside the |
391 | // elemental region is implicit, and later pass in lowering (stack |
392 | // reclaim) fir.do_loop will be in charge of emitting any stack |
393 | // save/restore if needed. |
394 | auto *bldr = &converter.getFirOpBuilder(); |
395 | mlir::Value sp = bldr->genStackSave(loc); |
396 | stmtCtx.attachCleanup( |
397 | [bldr, loc, sp]() { bldr->genStackRestore(loc, sp); }); |
398 | } |
399 | mlir::Value temp = |
400 | builder.createTemporary(loc, type, ".result", extents, resultLengths); |
401 | return toExtendedValue(loc, temp, extents, lengths); |
402 | }(); |
403 | |
404 | if (mustPopSymMap) |
405 | symMap.popScope(); |
406 | |
407 | // Place allocated result |
408 | if (allocatedResult) { |
409 | if (std::optional<Fortran::lower::CallInterface< |
410 | Fortran::lower::CallerInterface>::PassedEntity> |
411 | resultArg = caller.getPassedResult()) { |
412 | if (resultArg->passBy == PassBy::AddressAndLength) |
413 | caller.placeAddressAndLengthInput(*resultArg, |
414 | fir::getBase(*allocatedResult), |
415 | fir::getLen(*allocatedResult)); |
416 | else if (resultArg->passBy == PassBy::BaseAddress) |
417 | caller.placeInput(*resultArg, fir::getBase(*allocatedResult)); |
418 | else |
419 | fir::emitFatalError( |
420 | loc, "only expect character scalar result to be passed by ref"); |
421 | } |
422 | } |
423 | |
424 | // In older Fortran, procedure argument types are inferred. This may lead |
425 | // different view of what the function signature is in different locations. |
426 | // Casts are inserted as needed below to accommodate this. |
427 | |
428 | // The mlir::func::FuncOp type prevails, unless it has a different number of |
429 | // arguments which can happen in legal program if it was passed as a dummy |
430 | // procedure argument earlier with no further type information. |
431 | mlir::SymbolRefAttr funcSymbolAttr; |
432 | bool addHostAssociations = false; |
433 | if (!funcPointer) { |
434 | mlir::FunctionType funcOpType = caller.getFuncOp().getFunctionType(); |
435 | mlir::SymbolRefAttr symbolAttr = |
436 | builder.getSymbolRefAttr(caller.getMangledName()); |
437 | if (callSiteType.getNumResults() == funcOpType.getNumResults() && |
438 | callSiteType.getNumInputs() + 1 == funcOpType.getNumInputs() && |
439 | fir::anyFuncArgsHaveAttr(caller.getFuncOp(), |
440 | fir::getHostAssocAttrName())) { |
441 | // The number of arguments is off by one, and we're lowering a function |
442 | // with host associations. Modify call to include host associations |
443 | // argument by appending the value at the end of the operands. |
444 | assert(funcOpType.getInput(findHostAssocTuplePos(caller.getFuncOp())) == |
445 | converter.hostAssocTupleValue().getType()); |
446 | addHostAssociations = true; |
447 | } |
448 | // When this is not a call to an internal procedure (where there is a |
449 | // mismatch due to the extra argument, but the interface is otherwise |
450 | // explicit and safe), handle interface mismatch due to F77 implicit |
451 | // interface "abuse" with a function address cast if needed. |
452 | if (!addHostAssociations && |
453 | mustCastFuncOpToCopeWithImplicitInterfaceMismatch( |
454 | loc, converter, callSiteType, funcOpType)) |
455 | funcPointer = builder.create<fir::AddrOfOp>(loc, funcOpType, symbolAttr); |
456 | else |
457 | funcSymbolAttr = symbolAttr; |
458 | |
459 | // Issue a warning if the procedure name conflicts with |
460 | // a runtime function name a call to which has been already |
461 | // lowered (implying that the FuncOp has been created). |
462 | // The behavior is undefined in this case. |
463 | if (caller.getFuncOp()->hasAttrOfType<mlir::UnitAttr>( |
464 | fir::FIROpsDialect::getFirRuntimeAttrName())) |
465 | LLVM_DEBUG(mlir::emitWarning( |
466 | loc, |
467 | llvm::Twine("function name '") + |
468 | llvm::Twine(symbolAttr.getLeafReference()) + |
469 | llvm::Twine("' conflicts with a runtime function name used by " |
470 | "Flang - this may lead to undefined behavior"))); |
471 | } |
472 | |
473 | mlir::FunctionType funcType = |
474 | funcPointer ? callSiteType : caller.getFuncOp().getFunctionType(); |
475 | llvm::SmallVector<mlir::Value> operands; |
476 | // First operand of indirect call is the function pointer. Cast it to |
477 | // required function type for the call to handle procedures that have a |
478 | // compatible interface in Fortran, but that have different signatures in |
479 | // FIR. |
480 | if (funcPointer) { |
481 | operands.push_back( |
482 | mlir::isa<fir::BoxProcType>(funcPointer.getType()) |
483 | ? builder.create<fir::BoxAddrOp>(loc, funcType, funcPointer) |
484 | : builder.createConvert(loc, funcType, funcPointer)); |
485 | } |
486 | |
487 | // Deal with potential mismatches in arguments types. Passing an array to a |
488 | // scalar argument should for instance be tolerated here. |
489 | for (auto [fst, snd] : llvm::zip(caller.getInputs(), funcType.getInputs())) { |
490 | // When passing arguments to a procedure that can be called by implicit |
491 | // interface, allow any character actual arguments to be passed to dummy |
492 | // arguments of any type and vice versa. |
493 | mlir::Value cast; |
494 | auto *context = builder.getContext(); |
495 | if (mlir::isa<fir::BoxProcType>(snd) && |
496 | mlir::isa<mlir::FunctionType>(fst.getType())) { |
497 | auto funcTy = |
498 | mlir::FunctionType::get(context, std::nullopt, std::nullopt); |
499 | auto boxProcTy = builder.getBoxProcType(funcTy); |
500 | if (mlir::Value host = argumentHostAssocs(converter, fst)) { |
501 | cast = builder.create<fir::EmboxProcOp>( |
502 | loc, boxProcTy, llvm::ArrayRef<mlir::Value>{fst, host}); |
503 | } else { |
504 | cast = builder.create<fir::EmboxProcOp>(loc, boxProcTy, fst); |
505 | } |
506 | } else { |
507 | mlir::Type fromTy = fir::unwrapRefType(fst.getType()); |
508 | if (fir::isa_builtin_cptr_type(fromTy) && |
509 | Fortran::lower::isCPtrArgByValueType(snd)) { |
510 | cast = genRecordCPtrValueArg(builder, loc, fst, fromTy); |
511 | } else if (fir::isa_derived(snd) && !fir::isa_derived(fst.getType())) { |
512 | // TODO: remove this TODO once the old lowering is gone. |
513 | TODO(loc, "derived type argument passed by value"); |
514 | } else { |
515 | // With the lowering to HLFIR, box arguments have already been built |
516 | // according to the attributes, rank, bounds, and type they should have. |
517 | // Do not attempt any reboxing here that could break this. |
518 | bool legacyLowering = |
519 | !converter.getLoweringOptions().getLowerToHighLevelFIR(); |
520 | // When dealing with a dummy character argument (fir.boxchar), the |
521 | // effective argument might be a non-character raw pointer. This may |
522 | // happen when calling an implicit interface that was previously called |
523 | // with a character argument, or when calling an explicit interface with |
524 | // an IgnoreTKR dummy character arguments. Allow creating a fir.boxchar |
525 | // from the raw pointer, which requires a non-trivial type conversion. |
526 | const bool allowCharacterConversions = true; |
527 | bool isVolatile = fir::isa_volatile_type(snd); |
528 | cast = builder.createVolatileCast(loc, isVolatile, fst); |
529 | cast = builder.convertWithSemantics(loc, snd, cast, |
530 | allowCharacterConversions, |
531 | /*allowRebox=*/legacyLowering); |
532 | } |
533 | } |
534 | operands.push_back(cast); |
535 | } |
536 | |
537 | // Add host associations as necessary. |
538 | if (addHostAssociations) |
539 | operands.push_back(converter.hostAssocTupleValue()); |
540 | |
541 | mlir::Value callResult; |
542 | unsigned callNumResults; |
543 | fir::FortranProcedureFlagsEnumAttr procAttrs = |
544 | caller.getProcedureAttrs(builder.getContext()); |
545 | |
546 | if (!caller.getCallDescription().chevrons().empty()) { |
547 | // A call to a CUDA kernel with the chevron syntax. |
548 | |
549 | mlir::Type i32Ty = builder.getI32Type(); |
550 | mlir::Value one = builder.createIntegerConstant(loc, i32Ty, 1); |
551 | |
552 | mlir::Value grid_x, grid_y, grid_z; |
553 | if (caller.getCallDescription().chevrons()[0].GetType()->category() == |
554 | Fortran::common::TypeCategory::Integer) { |
555 | // If grid is an integer, it is converted to dim3(grid,1,1). Since z is |
556 | // not used for the number of thread blocks, it is omitted in the op. |
557 | grid_x = builder.createConvert( |
558 | loc, i32Ty, |
559 | fir::getBase(converter.genExprValue( |
560 | caller.getCallDescription().chevrons()[0], stmtCtx))); |
561 | grid_y = one; |
562 | grid_z = one; |
563 | } else { |
564 | auto dim3Addr = converter.genExprAddr( |
565 | caller.getCallDescription().chevrons()[0], stmtCtx); |
566 | grid_x = readDim3Value(builder, loc, fir::getBase(dim3Addr), "x"); |
567 | grid_y = readDim3Value(builder, loc, fir::getBase(dim3Addr), "y"); |
568 | grid_z = readDim3Value(builder, loc, fir::getBase(dim3Addr), "z"); |
569 | } |
570 | |
571 | mlir::Value block_x, block_y, block_z; |
572 | if (caller.getCallDescription().chevrons()[1].GetType()->category() == |
573 | Fortran::common::TypeCategory::Integer) { |
574 | // If block is an integer, it is converted to dim3(block,1,1). |
575 | block_x = builder.createConvert( |
576 | loc, i32Ty, |
577 | fir::getBase(converter.genExprValue( |
578 | caller.getCallDescription().chevrons()[1], stmtCtx))); |
579 | block_y = one; |
580 | block_z = one; |
581 | } else { |
582 | auto dim3Addr = converter.genExprAddr( |
583 | caller.getCallDescription().chevrons()[1], stmtCtx); |
584 | block_x = readDim3Value(builder, loc, fir::getBase(dim3Addr), "x"); |
585 | block_y = readDim3Value(builder, loc, fir::getBase(dim3Addr), "y"); |
586 | block_z = readDim3Value(builder, loc, fir::getBase(dim3Addr), "z"); |
587 | } |
588 | |
589 | mlir::Value bytes; // bytes is optional. |
590 | if (caller.getCallDescription().chevrons().size() > 2) |
591 | bytes = builder.createConvert( |
592 | loc, i32Ty, |
593 | fir::getBase(converter.genExprValue( |
594 | caller.getCallDescription().chevrons()[2], stmtCtx))); |
595 | |
596 | mlir::Value stream; // stream is optional. |
597 | if (caller.getCallDescription().chevrons().size() > 3) |
598 | stream = fir::getBase(converter.genExprAddr( |
599 | caller.getCallDescription().chevrons()[3], stmtCtx)); |
600 | |
601 | builder.create<cuf::KernelLaunchOp>( |
602 | loc, funcType.getResults(), funcSymbolAttr, grid_x, grid_y, grid_z, |
603 | block_x, block_y, block_z, bytes, stream, operands, |
604 | /*arg_attrs=*/nullptr, /*res_attrs=*/nullptr); |
605 | callNumResults = 0; |
606 | } else if (caller.requireDispatchCall()) { |
607 | // Procedure call requiring a dynamic dispatch. Call is created with |
608 | // fir.dispatch. |
609 | |
610 | // Get the raw procedure name. The procedure name is not mangled in the |
611 | // binding table, but there can be a suffix to distinguish bindings of |
612 | // the same name (which happens only when PRIVATE bindings exist in |
613 | // ancestor types in other modules). |
614 | const auto &ultimateSymbol = |
615 | caller.getCallDescription().proc().GetSymbol()->GetUltimate(); |
616 | std::string procName = ultimateSymbol.name().ToString(); |
617 | if (const auto &binding{ |
618 | ultimateSymbol.get<Fortran::semantics::ProcBindingDetails>()}; |
619 | binding.numPrivatesNotOverridden() > 0) |
620 | procName += "."s+ std::to_string(binding.numPrivatesNotOverridden()); |
621 | fir::DispatchOp dispatch; |
622 | if (std::optional<unsigned> passArg = caller.getPassArgIndex()) { |
623 | // PASS, PASS(arg-name) |
624 | // Note that caller.getInputs is used instead of operands to get the |
625 | // passed object because interface mismatch issues may have inserted a |
626 | // cast to the operand with a different declared type, which would break |
627 | // later type bound call resolution in the FIR to FIR pass. |
628 | dispatch = builder.create<fir::DispatchOp>( |
629 | loc, funcType.getResults(), builder.getStringAttr(procName), |
630 | caller.getInputs()[*passArg], operands, |
631 | builder.getI32IntegerAttr(*passArg), /*arg_attrs=*/nullptr, |
632 | /*res_attrs=*/nullptr, procAttrs); |
633 | } else { |
634 | // NOPASS |
635 | const Fortran::evaluate::Component *component = |
636 | caller.getCallDescription().proc().GetComponent(); |
637 | assert(component && "expect component for type-bound procedure call."); |
638 | |
639 | fir::ExtendedValue dataRefValue = Fortran::lower::convertDataRefToValue( |
640 | loc, converter, component->base(), symMap, stmtCtx); |
641 | mlir::Value passObject = fir::getBase(dataRefValue); |
642 | |
643 | if (fir::isa_ref_type(passObject.getType())) |
644 | passObject = builder.create<fir::LoadOp>(loc, passObject); |
645 | dispatch = builder.create<fir::DispatchOp>( |
646 | loc, funcType.getResults(), builder.getStringAttr(procName), |
647 | passObject, operands, nullptr, /*arg_attrs=*/nullptr, |
648 | /*res_attrs=*/nullptr, procAttrs); |
649 | } |
650 | callNumResults = dispatch.getNumResults(); |
651 | if (callNumResults != 0) |
652 | callResult = dispatch.getResult(0); |
653 | } else { |
654 | // Standard procedure call with fir.call. |
655 | auto call = builder.create<fir::CallOp>( |
656 | loc, funcType.getResults(), funcSymbolAttr, operands, |
657 | /*arg_attrs=*/nullptr, /*res_attrs=*/nullptr, procAttrs); |
658 | |
659 | callNumResults = call.getNumResults(); |
660 | if (callNumResults != 0) |
661 | callResult = call.getResult(0); |
662 | } |
663 | |
664 | std::optional<Fortran::evaluate::DynamicType> retTy = |
665 | caller.getCallDescription().proc().GetType(); |
666 | // With HLFIR lowering, isElemental must be set to true |
667 | // if we are producing an elemental call. In this case, |
668 | // the elemental results must not be destroyed, instead, |
669 | // the resulting array result will be finalized/destroyed |
670 | // as needed by hlfir.destroy. |
671 | const bool mustFinalizeResult = |
672 | !isElemental && callSiteType.getNumResults() > 0 && |
673 | !fir::isPointerType(callSiteType.getResult(0)) && retTy.has_value() && |
674 | (retTy->category() == Fortran::common::TypeCategory::Derived || |
675 | retTy->IsPolymorphic() || retTy->IsUnlimitedPolymorphic()); |
676 | |
677 | if (caller.mustSaveResult()) { |
678 | assert(allocatedResult.has_value()); |
679 | builder.create<fir::SaveResultOp>(loc, callResult, |
680 | fir::getBase(*allocatedResult), |
681 | arrayResultShape, resultLengths); |
682 | } |
683 | |
684 | if (evaluateInMemory) { |
685 | builder.setInsertionPointAfter(evaluateInMemory); |
686 | mlir::Value expr = evaluateInMemory.getResult(); |
687 | fir::FirOpBuilder *bldr = &converter.getFirOpBuilder(); |
688 | if (!isElemental) |
689 | stmtCtx.attachCleanup([bldr, loc, expr, mustFinalizeResult]() { |
690 | bldr->create<hlfir::DestroyOp>(loc, expr, |
691 | /*finalize=*/mustFinalizeResult); |
692 | }); |
693 | return {LoweredResult{hlfir::EntityWithAttributes{expr}}, |
694 | mustFinalizeResult}; |
695 | } |
696 | |
697 | if (allocatedResult) { |
698 | // The result must be optionally destroyed (if it is of a derived type |
699 | // that may need finalization or deallocation of the components). |
700 | // For an allocatable result we have to free the memory allocated |
701 | // for the top-level entity. Note that the Destroy calls below |
702 | // do not deallocate the top-level entity. The two clean-ups |
703 | // must be pushed in reverse order, so that the final order is: |
704 | // Destroy(desc) |
705 | // free(desc->base_addr) |
706 | allocatedResult->match( |
707 | [&](const fir::MutableBoxValue &box) { |
708 | if (box.isAllocatable()) { |
709 | // 9.7.3.2 point 4. Deallocate allocatable results. Note that |
710 | // finalization was done independently by calling |
711 | // genDerivedTypeDestroy above and is not triggered by this inline |
712 | // deallocation. |
713 | fir::FirOpBuilder *bldr = &converter.getFirOpBuilder(); |
714 | stmtCtx.attachCleanup([bldr, loc, box]() { |
715 | fir::factory::genFreememIfAllocated(*bldr, loc, box); |
716 | }); |
717 | } |
718 | }, |
719 | [](const auto &) {}); |
720 | |
721 | // 7.5.6.3 point 5. Derived-type finalization for nonpointer function. |
722 | bool resultIsFinalized = false; |
723 | // Check if the derived-type is finalizable if it is a monomorphic |
724 | // derived-type. |
725 | // For polymorphic and unlimited polymorphic enities call the runtime |
726 | // in any cases. |
727 | if (mustFinalizeResult) { |
728 | if (retTy->IsPolymorphic() || retTy->IsUnlimitedPolymorphic()) { |
729 | auto *bldr = &converter.getFirOpBuilder(); |
730 | stmtCtx.attachCleanup([bldr, loc, allocatedResult]() { |
731 | fir::runtime::genDerivedTypeDestroy(*bldr, loc, |
732 | fir::getBase(*allocatedResult)); |
733 | }); |
734 | resultIsFinalized = true; |
735 | } else { |
736 | const Fortran::semantics::DerivedTypeSpec &typeSpec = |
737 | retTy->GetDerivedTypeSpec(); |
738 | // If the result type may require finalization |
739 | // or have allocatable components, we need to make sure |
740 | // everything is properly finalized/deallocated. |
741 | if (Fortran::semantics::MayRequireFinalization(typeSpec) || |
742 | // We can use DerivedTypeDestroy even if finalization is not needed. |
743 | hlfir::mayHaveAllocatableComponent(funcType.getResults()[0])) { |
744 | auto *bldr = &converter.getFirOpBuilder(); |
745 | stmtCtx.attachCleanup([bldr, loc, allocatedResult]() { |
746 | mlir::Value box = bldr->createBox(loc, *allocatedResult); |
747 | fir::runtime::genDerivedTypeDestroy(*bldr, loc, box); |
748 | }); |
749 | resultIsFinalized = true; |
750 | } |
751 | } |
752 | } |
753 | return {LoweredResult{*allocatedResult}, resultIsFinalized}; |
754 | } |
755 | |
756 | // subroutine call |
757 | if (!resultType) |
758 | return {LoweredResult{fir::ExtendedValue{mlir::Value{}}}, |
759 | /*resultIsFinalized=*/false}; |
760 | |
761 | // For now, Fortran return values are implemented with a single MLIR |
762 | // function return value. |
763 | assert(callNumResults == 1 && "Expected exactly one result in FUNCTION call"); |
764 | (void)callNumResults; |
765 | |
766 | // Call a BIND(C) function that return a char. |
767 | if (caller.characterize().IsBindC() && |
768 | mlir::isa<fir::CharacterType>(funcType.getResults()[0])) { |
769 | fir::CharacterType charTy = |
770 | mlir::dyn_cast<fir::CharacterType>(funcType.getResults()[0]); |
771 | mlir::Value len = builder.createIntegerConstant( |
772 | loc, builder.getCharacterLengthType(), charTy.getLen()); |
773 | return { |
774 | LoweredResult{fir::ExtendedValue{fir::CharBoxValue{callResult, len}}}, |
775 | /*resultIsFinalized=*/false}; |
776 | } |
777 | |
778 | return {LoweredResult{fir::ExtendedValue{callResult}}, |
779 | /*resultIsFinalized=*/false}; |
780 | } |
781 | |
782 | static hlfir::EntityWithAttributes genStmtFunctionRef( |
783 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
784 | Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx, |
785 | const Fortran::evaluate::ProcedureRef &procRef) { |
786 | const Fortran::semantics::Symbol *symbol = procRef.proc().GetSymbol(); |
787 | assert(symbol && "expected symbol in ProcedureRef of statement functions"); |
788 | const auto &details = symbol->get<Fortran::semantics::SubprogramDetails>(); |
789 | fir::FirOpBuilder &builder = converter.getFirOpBuilder(); |
790 | |
791 | // Statement functions have their own scope, we just need to associate |
792 | // the dummy symbols to argument expressions. There are no |
793 | // optional/alternate return arguments. Statement functions cannot be |
794 | // recursive (directly or indirectly) so it is safe to add dummy symbols to |
795 | // the local map here. |
796 | symMap.pushScope(); |
797 | llvm::SmallVector<hlfir::AssociateOp> exprAssociations; |
798 | for (auto [arg, bind] : llvm::zip(details.dummyArgs(), procRef.arguments())) { |
799 | assert(arg && "alternate return in statement function"); |
800 | assert(bind && "optional argument in statement function"); |
801 | const auto *expr = bind->UnwrapExpr(); |
802 | // TODO: assumed type in statement function, that surprisingly seems |
803 | // allowed, probably because nobody thought of restricting this usage. |
804 | // gfortran/ifort compiles this. |
805 | assert(expr && "assumed type used as statement function argument"); |
806 | // As per Fortran 2018 C1580, statement function arguments can only be |
807 | // scalars. |
808 | // The only care is to use the dummy character explicit length if any |
809 | // instead of the actual argument length (that can be bigger). |
810 | hlfir::EntityWithAttributes loweredArg = Fortran::lower::convertExprToHLFIR( |
811 | loc, converter, *expr, symMap, stmtCtx); |
812 | fir::FortranVariableOpInterface variableIface = loweredArg.getIfVariable(); |
813 | if (!variableIface) { |
814 | // So far only FortranVariableOpInterface can be mapped to symbols. |
815 | // Create an hlfir.associate to create a variable from a potential |
816 | // value argument. |
817 | mlir::Type argType = converter.genType(*arg); |
818 | auto associate = hlfir::genAssociateExpr( |
819 | loc, builder, loweredArg, argType, toStringRef(arg->name())); |
820 | exprAssociations.push_back(associate); |
821 | variableIface = associate; |
822 | } |
823 | const Fortran::semantics::DeclTypeSpec *type = arg->GetType(); |
824 | if (type && |
825 | type->category() == Fortran::semantics::DeclTypeSpec::Character) { |
826 | // Instantiate character as if it was a normal dummy argument so that the |
827 | // statement function dummy character length is applied and dealt with |
828 | // correctly. |
829 | symMap.addSymbol(*arg, variableIface.getBase()); |
830 | Fortran::lower::mapSymbolAttributes(converter, *arg, symMap, stmtCtx); |
831 | } else { |
832 | // No need to create an extra hlfir.declare otherwise for |
833 | // numerical and logical scalar dummies. |
834 | symMap.addVariableDefinition(*arg, variableIface); |
835 | } |
836 | } |
837 | |
838 | // Explicitly map statement function host associated symbols to their |
839 | // parent scope lowered symbol box. |
840 | for (const Fortran::semantics::SymbolRef &sym : |
841 | Fortran::evaluate::CollectSymbols(*details.stmtFunction())) |
842 | if (const auto *details = |
843 | sym->detailsIf<Fortran::semantics::HostAssocDetails>()) |
844 | converter.copySymbolBinding(details->symbol(), sym); |
845 | |
846 | hlfir::Entity result = Fortran::lower::convertExprToHLFIR( |
847 | loc, converter, details.stmtFunction().value(), symMap, stmtCtx); |
848 | symMap.popScope(); |
849 | // The result must not be a variable. |
850 | result = hlfir::loadTrivialScalar(loc, builder, result); |
851 | if (result.isVariable()) |
852 | result = hlfir::Entity{builder.create<hlfir::AsExprOp>(loc, result)}; |
853 | for (auto associate : exprAssociations) |
854 | builder.create<hlfir::EndAssociateOp>(loc, associate); |
855 | return hlfir::EntityWithAttributes{result}; |
856 | } |
857 | |
858 | namespace { |
859 | // Structure to hold the information about the call and the lowering context. |
860 | // This structure is intended to help threading the information |
861 | // through the various lowering calls without having to pass every |
862 | // required structure one by one. |
863 | struct CallContext { |
864 | CallContext(const Fortran::evaluate::ProcedureRef &procRef, |
865 | std::optional<mlir::Type> resultType, mlir::Location loc, |
866 | Fortran::lower::AbstractConverter &converter, |
867 | Fortran::lower::SymMap &symMap, |
868 | Fortran::lower::StatementContext &stmtCtx) |
869 | : procRef{procRef}, converter{converter}, symMap{symMap}, |
870 | stmtCtx{stmtCtx}, resultType{resultType}, loc{loc} {} |
871 | |
872 | fir::FirOpBuilder &getBuilder() { return converter.getFirOpBuilder(); } |
873 | |
874 | std::string getProcedureName() const { |
875 | if (const Fortran::semantics::Symbol *sym = procRef.proc().GetSymbol()) |
876 | return sym->GetUltimate().name().ToString(); |
877 | return procRef.proc().GetName(); |
878 | } |
879 | |
880 | /// Is this a call to an elemental procedure with at least one array argument? |
881 | bool isElementalProcWithArrayArgs() const { |
882 | if (procRef.IsElemental()) |
883 | for (const std::optional<Fortran::evaluate::ActualArgument> &arg : |
884 | procRef.arguments()) |
885 | if (arg && arg->Rank() != 0) |
886 | return true; |
887 | return false; |
888 | } |
889 | |
890 | /// Is this a statement function reference? |
891 | bool isStatementFunctionCall() const { |
892 | if (const Fortran::semantics::Symbol *symbol = procRef.proc().GetSymbol()) |
893 | if (const auto *details = |
894 | symbol->detailsIf<Fortran::semantics::SubprogramDetails>()) |
895 | return details->stmtFunction().has_value(); |
896 | return false; |
897 | } |
898 | |
899 | /// Is this a call to a BIND(C) procedure? |
900 | bool isBindcCall() const { |
901 | if (const Fortran::semantics::Symbol *symbol = procRef.proc().GetSymbol()) |
902 | return Fortran::semantics::IsBindCProcedure(*symbol); |
903 | return false; |
904 | } |
905 | |
906 | const Fortran::evaluate::ProcedureRef &procRef; |
907 | Fortran::lower::AbstractConverter &converter; |
908 | Fortran::lower::SymMap &symMap; |
909 | Fortran::lower::StatementContext &stmtCtx; |
910 | std::optional<mlir::Type> resultType; |
911 | mlir::Location loc; |
912 | }; |
913 | |
914 | using ExvAndCleanup = |
915 | std::pair<fir::ExtendedValue, std::optional<hlfir::CleanupFunction>>; |
916 | } // namespace |
917 | |
918 | // Helper to transform a fir::ExtendedValue to an hlfir::EntityWithAttributes. |
919 | static hlfir::EntityWithAttributes |
920 | extendedValueToHlfirEntity(mlir::Location loc, fir::FirOpBuilder &builder, |
921 | const fir::ExtendedValue &exv, |
922 | llvm::StringRef name) { |
923 | mlir::Value firBase = fir::getBase(exv); |
924 | mlir::Type firBaseTy = firBase.getType(); |
925 | if (fir::isa_trivial(firBaseTy)) |
926 | return hlfir::EntityWithAttributes{firBase}; |
927 | if (auto charTy = mlir::dyn_cast<fir::CharacterType>(firBase.getType())) { |
928 | // CHAR() intrinsic and BIND(C) procedures returning CHARACTER(1) |
929 | // are lowered to a fir.char<kind,1> that is not in memory. |
930 | // This tends to cause a lot of bugs because the rest of the |
931 | // infrastructure is mostly tested with characters that are |
932 | // in memory. |
933 | // To avoid having to deal with this special case here and there, |
934 | // place it in memory here. If this turns out to be suboptimal, |
935 | // this could be fixed, but for now llvm opt -O1 is able to get |
936 | // rid of the memory indirection in a = char(b), so there is |
937 | // little incentive to increase the compiler complexity. |
938 | hlfir::Entity storage{builder.createTemporary(loc, charTy)}; |
939 | builder.create<fir::StoreOp>(loc, firBase, storage); |
940 | auto asExpr = builder.create<hlfir::AsExprOp>( |
941 | loc, storage, /*mustFree=*/builder.createBool(loc, false)); |
942 | return hlfir::EntityWithAttributes{asExpr.getResult()}; |
943 | } |
944 | return hlfir::genDeclare(loc, builder, exv, name, |
945 | fir::FortranVariableFlagsAttr{}); |
946 | } |
947 | namespace { |
948 | /// Structure to hold the clean-up related to a dummy argument preparation |
949 | /// that may have to be done after a call (copy-out or temporary deallocation). |
950 | struct CallCleanUp { |
951 | struct CopyIn { |
952 | void genCleanUp(mlir::Location loc, fir::FirOpBuilder &builder) { |
953 | builder.create<hlfir::CopyOutOp>(loc, tempBox, wasCopied, copyBackVar); |
954 | } |
955 | // address of the descriptor holding the temp if a temp was created. |
956 | mlir::Value tempBox; |
957 | // Boolean indicating if a copy was made or not. |
958 | mlir::Value wasCopied; |
959 | // copyBackVar may be null if copy back is not needed. |
960 | mlir::Value copyBackVar; |
961 | }; |
962 | struct ExprAssociate { |
963 | void genCleanUp(mlir::Location loc, fir::FirOpBuilder &builder) { |
964 | builder.create<hlfir::EndAssociateOp>(loc, tempVar, mustFree); |
965 | } |
966 | mlir::Value tempVar; |
967 | mlir::Value mustFree; |
968 | }; |
969 | |
970 | /// Generate clean-up code. |
971 | /// If \p postponeAssociates is true, the ExprAssociate clean-up |
972 | /// is not generated, and instead the corresponding CallCleanUp |
973 | /// object is returned as the result. |
974 | std::optional<CallCleanUp> genCleanUp(mlir::Location loc, |
975 | fir::FirOpBuilder &builder, |
976 | bool postponeAssociates) { |
977 | std::optional<CallCleanUp> postponed; |
978 | Fortran::common::visit(Fortran::common::visitors{ |
979 | [&](CopyIn &c) { c.genCleanUp(loc, builder); }, |
980 | [&](ExprAssociate &c) { |
981 | if (postponeAssociates) |
982 | postponed = CallCleanUp{c}; |
983 | else |
984 | c.genCleanUp(loc, builder); |
985 | }, |
986 | }, |
987 | cleanUp); |
988 | return postponed; |
989 | } |
990 | std::variant<CopyIn, ExprAssociate> cleanUp; |
991 | }; |
992 | |
993 | /// Structure representing a prepared dummy argument. |
994 | /// It holds the value to be passed in the call and any related |
995 | /// clean-ups to be done after the call. |
996 | struct PreparedDummyArgument { |
997 | void pushCopyInCleanUp(mlir::Value tempBox, mlir::Value wasCopied, |
998 | mlir::Value copyBackVar) { |
999 | cleanups.emplace_back( |
1000 | Args: CallCleanUp{CallCleanUp::CopyIn{tempBox, wasCopied, copyBackVar}}); |
1001 | } |
1002 | void pushExprAssociateCleanUp(mlir::Value tempVar, mlir::Value wasCopied) { |
1003 | cleanups.emplace_back( |
1004 | Args: CallCleanUp{CallCleanUp::ExprAssociate{tempVar, wasCopied}}); |
1005 | } |
1006 | void pushExprAssociateCleanUp(hlfir::AssociateOp associate) { |
1007 | mlir::Value hlfirBase = associate.getBase(); |
1008 | mlir::Value firBase = associate.getFirBase(); |
1009 | cleanups.emplace_back(CallCleanUp{CallCleanUp::ExprAssociate{ |
1010 | hlfir::mayHaveAllocatableComponent(hlfirBase.getType()) ? hlfirBase |
1011 | : firBase, |
1012 | associate.getMustFreeStrorageFlag()}}); |
1013 | } |
1014 | |
1015 | mlir::Value dummy; |
1016 | // NOTE: the clean-ups are executed in reverse order. |
1017 | llvm::SmallVector<CallCleanUp, 2> cleanups; |
1018 | }; |
1019 | |
1020 | /// Structure to help conditionally preparing a dummy argument based |
1021 | /// on the actual argument presence. |
1022 | /// It helps "wrapping" the dummy and the clean-up information in |
1023 | /// an if (present) {...}: |
1024 | /// |
1025 | /// %conditionallyPrepared = fir.if (%present) { |
1026 | /// fir.result %preparedDummy |
1027 | /// } else { |
1028 | /// fir.result %absent |
1029 | /// } |
1030 | /// |
1031 | struct ConditionallyPreparedDummy { |
1032 | /// Create ConditionallyPreparedDummy from a preparedDummy that must |
1033 | /// be wrapped in a fir.if. |
1034 | ConditionallyPreparedDummy(PreparedDummyArgument &preparedDummy) { |
1035 | thenResultValues.push_back(preparedDummy.dummy); |
1036 | for (const CallCleanUp &c : preparedDummy.cleanups) { |
1037 | if (const auto *copyInCleanUp = |
1038 | std::get_if<CallCleanUp::CopyIn>(&c.cleanUp)) { |
1039 | thenResultValues.push_back(copyInCleanUp->wasCopied); |
1040 | if (copyInCleanUp->copyBackVar) |
1041 | thenResultValues.push_back(copyInCleanUp->copyBackVar); |
1042 | } else { |
1043 | const auto &exprAssociate = |
1044 | std::get<CallCleanUp::ExprAssociate>(c.cleanUp); |
1045 | thenResultValues.push_back(exprAssociate.tempVar); |
1046 | thenResultValues.push_back(exprAssociate.mustFree); |
1047 | } |
1048 | } |
1049 | } |
1050 | |
1051 | /// Get the result types of the wrapping fir.if that must be created. |
1052 | llvm::SmallVector<mlir::Type> getIfResulTypes() const { |
1053 | llvm::SmallVector<mlir::Type> types; |
1054 | for (mlir::Value res : thenResultValues) |
1055 | types.push_back(res.getType()); |
1056 | return types; |
1057 | } |
1058 | |
1059 | /// Generate the "fir.result %preparedDummy" in the then branch of the |
1060 | /// wrapping fir.if. |
1061 | void genThenResult(mlir::Location loc, fir::FirOpBuilder &builder) const { |
1062 | builder.create<fir::ResultOp>(loc, thenResultValues); |
1063 | } |
1064 | |
1065 | /// Generate the "fir.result %absent" in the else branch of the |
1066 | /// wrapping fir.if. |
1067 | void genElseResult(mlir::Location loc, fir::FirOpBuilder &builder) const { |
1068 | llvm::SmallVector<mlir::Value> elseResultValues; |
1069 | mlir::Type i1Type = builder.getI1Type(); |
1070 | for (mlir::Value res : thenResultValues) { |
1071 | mlir::Type type = res.getType(); |
1072 | if (type == i1Type) |
1073 | elseResultValues.push_back(builder.createBool(loc, false)); |
1074 | else |
1075 | elseResultValues.push_back(builder.genAbsentOp(loc, type)); |
1076 | } |
1077 | builder.create<fir::ResultOp>(loc, elseResultValues); |
1078 | } |
1079 | |
1080 | /// Once the fir.if has been created, get the resulting %conditionallyPrepared |
1081 | /// dummy argument. |
1082 | PreparedDummyArgument |
1083 | getPreparedDummy(fir::IfOp ifOp, |
1084 | const PreparedDummyArgument &unconditionalDummy) { |
1085 | PreparedDummyArgument preparedDummy; |
1086 | preparedDummy.dummy = ifOp.getResults()[0]; |
1087 | for (const CallCleanUp &c : unconditionalDummy.cleanups) { |
1088 | if (const auto *copyInCleanUp = |
1089 | std::get_if<CallCleanUp::CopyIn>(&c.cleanUp)) { |
1090 | mlir::Value copyBackVar; |
1091 | if (copyInCleanUp->copyBackVar) |
1092 | copyBackVar = ifOp.getResults().back(); |
1093 | // tempBox is an hlfir.copy_in argument created outside of the |
1094 | // fir.if region. It needs not to be threaded as a fir.if result. |
1095 | preparedDummy.pushCopyInCleanUp(copyInCleanUp->tempBox, |
1096 | ifOp.getResults()[1], copyBackVar); |
1097 | } else { |
1098 | preparedDummy.pushExprAssociateCleanUp(ifOp.getResults()[1], |
1099 | ifOp.getResults()[2]); |
1100 | } |
1101 | } |
1102 | return preparedDummy; |
1103 | } |
1104 | |
1105 | llvm::SmallVector<mlir::Value> thenResultValues; |
1106 | }; |
1107 | } // namespace |
1108 | |
1109 | /// Fix-up the fact that it is supported to pass a character procedure |
1110 | /// designator to a non character procedure dummy procedure and vice-versa, even |
1111 | /// in case of explicit interface. Uglier cases where an object is passed as |
1112 | /// procedure designator or vice versa are handled only for implicit interfaces |
1113 | /// (refused by semantics with explicit interface), and handled with a funcOp |
1114 | /// cast like other implicit interface mismatches. |
1115 | static hlfir::Entity fixProcedureDummyMismatch(mlir::Location loc, |
1116 | fir::FirOpBuilder &builder, |
1117 | hlfir::Entity actual, |
1118 | mlir::Type dummyType) { |
1119 | if (mlir::isa<fir::BoxProcType>(actual.getType()) && |
1120 | fir::isCharacterProcedureTuple(dummyType)) { |
1121 | mlir::Value length = |
1122 | builder.create<fir::UndefOp>(loc, builder.getCharacterLengthType()); |
1123 | mlir::Value tuple = fir::factory::createCharacterProcedureTuple( |
1124 | builder, loc, dummyType, actual, length); |
1125 | return hlfir::Entity{tuple}; |
1126 | } |
1127 | assert(fir::isCharacterProcedureTuple(actual.getType()) && |
1128 | mlir::isa<fir::BoxProcType>(dummyType) && |
1129 | "unsupported dummy procedure mismatch with the actual argument"); |
1130 | mlir::Value boxProc = fir::factory::extractCharacterProcedureTuple( |
1131 | builder, loc, actual, /*openBoxProc=*/false) |
1132 | .first; |
1133 | return hlfir::Entity{boxProc}; |
1134 | } |
1135 | |
1136 | mlir::Value static getZeroLowerBounds(mlir::Location loc, |
1137 | fir::FirOpBuilder &builder, |
1138 | hlfir::Entity entity) { |
1139 | assert(!entity.isAssumedRank() && |
1140 | "assumed-rank must use fir.rebox_assumed_rank"); |
1141 | if (entity.getRank() < 1) |
1142 | return {}; |
1143 | mlir::Value zero = |
1144 | builder.createIntegerConstant(loc, builder.getIndexType(), 0); |
1145 | llvm::SmallVector<mlir::Value> lowerBounds(entity.getRank(), zero); |
1146 | return builder.genShift(loc, lowerBounds); |
1147 | } |
1148 | |
1149 | static bool |
1150 | isSimplyContiguous(const Fortran::evaluate::ActualArgument &arg, |
1151 | Fortran::evaluate::FoldingContext &foldingContext) { |
1152 | if (const auto *expr = arg.UnwrapExpr()) |
1153 | return Fortran::evaluate::IsSimplyContiguous(*expr, foldingContext); |
1154 | const Fortran::semantics::Symbol *sym = arg.GetAssumedTypeDummy(); |
1155 | assert(sym && |
1156 | "expect ActualArguments to be expression or assumed-type symbols"); |
1157 | return sym->Rank() == 0 || |
1158 | Fortran::evaluate::IsSimplyContiguous(*sym, foldingContext); |
1159 | } |
1160 | |
1161 | static bool isParameterObjectOrSubObject(hlfir::Entity entity) { |
1162 | mlir::Value base = entity; |
1163 | bool foundParameter = false; |
1164 | while (mlir::Operation *op = base ? base.getDefiningOp() : nullptr) { |
1165 | base = |
1166 | llvm::TypeSwitch<mlir::Operation *, mlir::Value>(op) |
1167 | .Case<hlfir::DeclareOp>([&](auto declare) -> mlir::Value { |
1168 | foundParameter |= hlfir::Entity{declare}.isParameter(); |
1169 | return foundParameter ? mlir::Value{} : declare.getMemref(); |
1170 | }) |
1171 | .Case<hlfir::DesignateOp, hlfir::ParentComponentOp, fir::EmboxOp>( |
1172 | [&](auto op) -> mlir::Value { return op.getMemref(); }) |
1173 | .Case<fir::ReboxOp>( |
1174 | [&](auto rebox) -> mlir::Value { return rebox.getBox(); }) |
1175 | .Case<fir::ConvertOp>( |
1176 | [&](auto convert) -> mlir::Value { return convert.getValue(); }) |
1177 | .Default([](mlir::Operation *) -> mlir::Value { return nullptr; }); |
1178 | } |
1179 | return foundParameter; |
1180 | } |
1181 | |
1182 | /// When dummy is not ALLOCATABLE, POINTER and is not passed in register, |
1183 | /// prepare the actual argument according to the interface. Do as needed: |
1184 | /// - address element if this is an array argument in an elemental call. |
1185 | /// - set dynamic type to the dummy type if the dummy is not polymorphic. |
1186 | /// - copy-in into contiguous variable if the dummy must be contiguous |
1187 | /// - copy into a temporary if the dummy has the VALUE attribute. |
1188 | /// - package the prepared dummy as required (fir.box, fir.class, |
1189 | /// fir.box_char...). |
1190 | /// This function should only be called with an actual that is present. |
1191 | /// The optional aspects must be handled by this function user. |
1192 | static PreparedDummyArgument preparePresentUserCallActualArgument( |
1193 | mlir::Location loc, fir::FirOpBuilder &builder, |
1194 | const Fortran::lower::PreparedActualArgument &preparedActual, |
1195 | mlir::Type dummyType, |
1196 | const Fortran::lower::CallerInterface::PassedEntity &arg, |
1197 | CallContext &callContext) { |
1198 | |
1199 | Fortran::evaluate::FoldingContext &foldingContext = |
1200 | callContext.converter.getFoldingContext(); |
1201 | |
1202 | // Step 1: get the actual argument, which includes addressing the |
1203 | // element if this is an array in an elemental call. |
1204 | hlfir::Entity actual = preparedActual.getActual(loc, builder); |
1205 | |
1206 | // Handle procedure arguments (procedure pointers should go through |
1207 | // prepareProcedurePointerActualArgument). |
1208 | if (hlfir::isFortranProcedureValue(dummyType)) { |
1209 | // Procedure pointer or function returns procedure pointer actual to |
1210 | // procedure dummy. |
1211 | if (actual.isProcedurePointer()) { |
1212 | actual = hlfir::derefPointersAndAllocatables(loc, builder, actual); |
1213 | return PreparedDummyArgument{actual, /*cleanups=*/{}}; |
1214 | } |
1215 | // Procedure actual to procedure dummy. |
1216 | assert(actual.isProcedure()); |
1217 | // Do nothing if this is a procedure argument. It is already a |
1218 | // fir.boxproc/fir.tuple<fir.boxproc, len> as it should. |
1219 | if (!mlir::isa<fir::BoxProcType>(actual.getType()) && |
1220 | actual.getType() != dummyType) |
1221 | // The actual argument may be a procedure that returns character (a |
1222 | // fir.tuple<fir.boxproc, len>) while the dummy is not. Extract the tuple |
1223 | // in that case. |
1224 | actual = fixProcedureDummyMismatch(loc, builder, actual, dummyType); |
1225 | return PreparedDummyArgument{actual, /*cleanups=*/{}}; |
1226 | } |
1227 | |
1228 | const bool ignoreTKRtype = arg.testTKR(Fortran::common::IgnoreTKR::Type); |
1229 | const bool passingPolymorphicToNonPolymorphic = |
1230 | actual.isPolymorphic() && !fir::isPolymorphicType(dummyType) && |
1231 | !ignoreTKRtype; |
1232 | |
1233 | // When passing a CLASS(T) to TYPE(T), only the "T" part must be |
1234 | // passed. Unless the entity is a scalar passed by raw address, a |
1235 | // new descriptor must be made using the dummy argument type as |
1236 | // dynamic type. This must be done before any copy/copy-in because the |
1237 | // dynamic type matters to determine the contiguity. |
1238 | const bool mustSetDynamicTypeToDummyType = |
1239 | passingPolymorphicToNonPolymorphic && |
1240 | (actual.isArray() || mlir::isa<fir::BaseBoxType>(dummyType)); |
1241 | |
1242 | // The simple contiguity of the actual is "lost" when passing a polymorphic |
1243 | // to a non polymorphic entity because the dummy dynamic type matters for |
1244 | // the contiguity. |
1245 | const bool mustDoCopyInOut = |
1246 | actual.isArray() && arg.mustBeMadeContiguous() && |
1247 | (passingPolymorphicToNonPolymorphic || |
1248 | !isSimplyContiguous(*arg.entity, foldingContext)); |
1249 | |
1250 | const bool actualIsAssumedRank = actual.isAssumedRank(); |
1251 | // Create dummy type with actual argument rank when the dummy is an assumed |
1252 | // rank. That way, all the operation to create dummy descriptors are ranked if |
1253 | // the actual argument is ranked, which allows simple code generation. |
1254 | // Also do the same when the dummy is a sequence associated descriptor |
1255 | // because the actual shape/rank may mismatch with the dummy, and the dummy |
1256 | // may be an assumed-size array, so any descriptor manipulation should use the |
1257 | // actual argument shape information. A descriptor with the dummy shape |
1258 | // information will be created later when all actual arguments are ready. |
1259 | mlir::Type dummyTypeWithActualRank = dummyType; |
1260 | if (auto baseBoxDummy = mlir::dyn_cast<fir::BaseBoxType>(dummyType)) { |
1261 | if (baseBoxDummy.isAssumedRank() || |
1262 | arg.testTKR(Fortran::common::IgnoreTKR::Rank) || |
1263 | arg.isSequenceAssociatedDescriptor()) { |
1264 | mlir::Type actualTy = |
1265 | hlfir::getFortranElementOrSequenceType(actual.getType()); |
1266 | dummyTypeWithActualRank = baseBoxDummy.getBoxTypeWithNewShape(actualTy); |
1267 | } |
1268 | } |
1269 | // Preserve the actual type in the argument preparation in case IgnoreTKR(t) |
1270 | // is set (descriptors must be created with the actual type in this case, and |
1271 | // copy-in/copy-out should be driven by the contiguity with regard to the |
1272 | // actual type). |
1273 | if (ignoreTKRtype) { |
1274 | if (auto boxCharType = |
1275 | mlir::dyn_cast<fir::BoxCharType>(dummyTypeWithActualRank)) { |
1276 | auto maybeActualCharType = |
1277 | mlir::dyn_cast<fir::CharacterType>(actual.getFortranElementType()); |
1278 | if (!maybeActualCharType || |
1279 | maybeActualCharType.getFKind() != boxCharType.getKind()) { |
1280 | // When passing to a fir.boxchar with ignore(tk), prepare the argument |
1281 | // as if only the raw address must be passed. |
1282 | dummyTypeWithActualRank = |
1283 | fir::ReferenceType::get(actual.getElementOrSequenceType()); |
1284 | } |
1285 | // Otherwise, the actual is already a character with the same kind as the |
1286 | // dummy and can be passed normally. |
1287 | } else { |
1288 | dummyTypeWithActualRank = fir::changeElementType( |
1289 | dummyTypeWithActualRank, actual.getFortranElementType(), |
1290 | actual.isPolymorphic()); |
1291 | } |
1292 | } |
1293 | |
1294 | PreparedDummyArgument preparedDummy; |
1295 | |
1296 | // Helpers to generate hlfir.copy_in operation and register the related |
1297 | // hlfir.copy_out creation. |
1298 | auto genCopyIn = [&](hlfir::Entity var, bool doCopyOut) -> hlfir::Entity { |
1299 | auto baseBoxTy = mlir::dyn_cast<fir::BaseBoxType>(var.getType()); |
1300 | assert(baseBoxTy && "expect non simply contiguous variables to be boxes"); |
1301 | // Create allocatable descriptor for the potential temporary. |
1302 | mlir::Type tempBoxType = baseBoxTy.getBoxTypeWithNewAttr( |
1303 | fir::BaseBoxType::Attribute::Allocatable); |
1304 | mlir::Value tempBox = builder.createTemporary(loc, tempBoxType); |
1305 | auto copyIn = builder.create<hlfir::CopyInOp>( |
1306 | loc, var, tempBox, /*var_is_present=*/mlir::Value{}); |
1307 | // Register the copy-out after the call. |
1308 | preparedDummy.pushCopyInCleanUp(copyIn.getTempBox(), copyIn.getWasCopied(), |
1309 | doCopyOut ? copyIn.getVar() |
1310 | : mlir::Value{}); |
1311 | return hlfir::Entity{copyIn.getCopiedIn()}; |
1312 | }; |
1313 | |
1314 | auto genSetDynamicTypeToDummyType = [&](hlfir::Entity var) -> hlfir::Entity { |
1315 | fir::BaseBoxType boxType = fir::BoxType::get( |
1316 | hlfir::getFortranElementOrSequenceType(dummyTypeWithActualRank)); |
1317 | if (actualIsAssumedRank) |
1318 | return hlfir::Entity{builder.create<fir::ReboxAssumedRankOp>( |
1319 | loc, boxType, var, fir::LowerBoundModifierAttribute::SetToOnes)}; |
1320 | // Use actual shape when creating descriptor with dummy type, the dummy |
1321 | // shape may be unknown in case of sequence association. |
1322 | mlir::Type actualTy = |
1323 | hlfir::getFortranElementOrSequenceType(actual.getType()); |
1324 | boxType = boxType.getBoxTypeWithNewShape(actualTy); |
1325 | return hlfir::Entity{builder.create<fir::ReboxOp>(loc, boxType, var, |
1326 | /*shape=*/mlir::Value{}, |
1327 | /*slice=*/mlir::Value{})}; |
1328 | }; |
1329 | |
1330 | // Step 2: prepare the storage for the dummy arguments, ensuring that it |
1331 | // matches the dummy requirements (e.g., must be contiguous or must be |
1332 | // a temporary). |
1333 | hlfir::Entity entity = |
1334 | hlfir::derefPointersAndAllocatables(loc, builder, actual); |
1335 | if (entity.isVariable()) { |
1336 | // Set dynamic type if needed before any copy-in or copy so that the dummy |
1337 | // is contiguous according to the dummy type. |
1338 | if (mustSetDynamicTypeToDummyType) |
1339 | entity = genSetDynamicTypeToDummyType(entity); |
1340 | if (arg.hasValueAttribute() || |
1341 | // Constant expressions might be lowered as variables with |
1342 | // 'parameter' attribute. Even though the constant expressions |
1343 | // are not definable and explicit assignments to them are not |
1344 | // possible, we have to create a temporary copies when we pass |
1345 | // them down the call stack because of potential compiler |
1346 | // generated writes in copy-out. |
1347 | isParameterObjectOrSubObject(entity)) { |
1348 | // Make a copy in a temporary. |
1349 | auto copy = builder.create<hlfir::AsExprOp>(loc, entity); |
1350 | mlir::Type storageType = entity.getType(); |
1351 | mlir::NamedAttribute byRefAttr = fir::getAdaptToByRefAttr(builder); |
1352 | hlfir::AssociateOp associate = hlfir::genAssociateExpr( |
1353 | loc, builder, hlfir::Entity{copy}, storageType, "", byRefAttr); |
1354 | entity = hlfir::Entity{associate.getBase()}; |
1355 | // Register the temporary destruction after the call. |
1356 | preparedDummy.pushExprAssociateCleanUp(associate); |
1357 | } else if (mustDoCopyInOut) { |
1358 | // Copy-in non contiguous variables. |
1359 | // TODO: for non-finalizable monomorphic derived type actual |
1360 | // arguments associated with INTENT(OUT) dummy arguments |
1361 | // we may avoid doing the copy and only allocate the temporary. |
1362 | // The codegen would do a "mold" allocation instead of "sourced" |
1363 | // allocation for the temp in this case. We can communicate |
1364 | // this to the codegen via some CopyInOp flag. |
1365 | // This is a performance concern. |
1366 | entity = genCopyIn(entity, arg.mayBeModifiedByCall()); |
1367 | } |
1368 | } else { |
1369 | const Fortran::lower::SomeExpr *expr = arg.entity->UnwrapExpr(); |
1370 | assert(expr && "expression actual argument cannot be an assumed type"); |
1371 | // The actual is an expression value, place it into a temporary |
1372 | // and register the temporary destruction after the call. |
1373 | mlir::Type storageType = callContext.converter.genType(*expr); |
1374 | mlir::NamedAttribute byRefAttr = fir::getAdaptToByRefAttr(builder); |
1375 | hlfir::AssociateOp associate = hlfir::genAssociateExpr( |
1376 | loc, builder, entity, storageType, "", byRefAttr); |
1377 | entity = hlfir::Entity{associate.getBase()}; |
1378 | preparedDummy.pushExprAssociateCleanUp(associate); |
1379 | // Rebox the actual argument to the dummy argument's type, and make sure |
1380 | // that we pass a contiguous entity (i.e. make copy-in, if needed). |
1381 | // |
1382 | // TODO: this can probably be optimized by associating the expression with |
1383 | // properly typed temporary, but this needs either a new operation or |
1384 | // making the hlfir.associate more complex. |
1385 | if (mustSetDynamicTypeToDummyType) { |
1386 | entity = genSetDynamicTypeToDummyType(entity); |
1387 | entity = genCopyIn(entity, /*doCopyOut=*/false); |
1388 | } |
1389 | } |
1390 | |
1391 | // Step 3: now that the dummy argument storage has been prepared, package |
1392 | // it according to the interface. |
1393 | mlir::Value addr; |
1394 | if (mlir::isa<fir::BoxCharType>(dummyTypeWithActualRank)) { |
1395 | // Cast the argument to match the volatility of the dummy argument. |
1396 | auto nonVolatileEntity = hlfir::Entity{builder.createVolatileCast( |
1397 | loc, fir::isa_volatile_type(dummyType), entity)}; |
1398 | addr = hlfir::genVariableBoxChar(loc, builder, nonVolatileEntity); |
1399 | } else if (mlir::isa<fir::BaseBoxType>(dummyTypeWithActualRank)) { |
1400 | entity = hlfir::genVariableBox(loc, builder, entity); |
1401 | // Ensures the box has the right attributes and that it holds an |
1402 | // addendum if needed. |
1403 | fir::BaseBoxType actualBoxType = |
1404 | mlir::cast<fir::BaseBoxType>(entity.getType()); |
1405 | mlir::Type boxEleType = actualBoxType.getEleTy(); |
1406 | // For now, assume it is not OK to pass the allocatable/pointer |
1407 | // descriptor to a non pointer/allocatable dummy. That is a strict |
1408 | // interpretation of 18.3.6 point 4 that stipulates the descriptor |
1409 | // has the dummy attributes in BIND(C) contexts. |
1410 | const bool actualBoxHasAllocatableOrPointerFlag = |
1411 | fir::isa_ref_type(boxEleType); |
1412 | // Fortran 2018 18.5.3, pp3: BIND(C) non pointer allocatable descriptors |
1413 | // must have zero lower bounds. |
1414 | bool needsZeroLowerBounds = callContext.isBindcCall() && entity.isArray(); |
1415 | // On the callee side, the current code generated for unlimited |
1416 | // polymorphic might unconditionally read the addendum. Intrinsic type |
1417 | // descriptors may not have an addendum, the rebox below will create a |
1418 | // descriptor with an addendum in such case. |
1419 | const bool actualBoxHasAddendum = fir::boxHasAddendum(actualBoxType); |
1420 | const bool needToAddAddendum = |
1421 | fir::isUnlimitedPolymorphicType(dummyTypeWithActualRank) && |
1422 | !actualBoxHasAddendum; |
1423 | if (needToAddAddendum || actualBoxHasAllocatableOrPointerFlag || |
1424 | needsZeroLowerBounds) { |
1425 | if (actualIsAssumedRank) { |
1426 | auto lbModifier = needsZeroLowerBounds |
1427 | ? fir::LowerBoundModifierAttribute::SetToZeroes |
1428 | : fir::LowerBoundModifierAttribute::SetToOnes; |
1429 | entity = hlfir::Entity{builder.create<fir::ReboxAssumedRankOp>( |
1430 | loc, dummyTypeWithActualRank, entity, lbModifier)}; |
1431 | } else { |
1432 | mlir::Value shift{}; |
1433 | if (needsZeroLowerBounds) |
1434 | shift = getZeroLowerBounds(loc, builder, entity); |
1435 | entity = hlfir::Entity{builder.create<fir::ReboxOp>( |
1436 | loc, dummyTypeWithActualRank, entity, /*shape=*/shift, |
1437 | /*slice=*/mlir::Value{})}; |
1438 | } |
1439 | } |
1440 | addr = entity; |
1441 | } else { |
1442 | addr = hlfir::genVariableRawAddress(loc, builder, entity); |
1443 | } |
1444 | |
1445 | // If the volatility of the input type does not match the dummy type, |
1446 | // we need to cast the argument. |
1447 | const bool isToTypeVolatile = fir::isa_volatile_type(dummyTypeWithActualRank); |
1448 | addr = builder.createVolatileCast(loc, isToTypeVolatile, addr); |
1449 | |
1450 | // For ranked actual passed to assumed-rank dummy, the cast to assumed-rank |
1451 | // box is inserted when building the fir.call op. Inserting it here would |
1452 | // cause the fir.if results to be assumed-rank in case of OPTIONAL dummy, |
1453 | // causing extra runtime costs due to the unknown runtime size of assumed-rank |
1454 | // descriptors. |
1455 | // For TKR dummy characters, the boxchar creation also happens later when |
1456 | // creating the fir.call . |
1457 | preparedDummy.dummy = |
1458 | builder.createConvert(loc, dummyTypeWithActualRank, addr); |
1459 | return preparedDummy; |
1460 | } |
1461 | |
1462 | /// When dummy is not ALLOCATABLE, POINTER and is not passed in register, |
1463 | /// prepare the actual argument according to the interface, taking care |
1464 | /// of any optional aspect. |
1465 | static PreparedDummyArgument prepareUserCallActualArgument( |
1466 | mlir::Location loc, fir::FirOpBuilder &builder, |
1467 | const Fortran::lower::PreparedActualArgument &preparedActual, |
1468 | mlir::Type dummyType, |
1469 | const Fortran::lower::CallerInterface::PassedEntity &arg, |
1470 | CallContext &callContext) { |
1471 | if (!preparedActual.handleDynamicOptional()) |
1472 | return preparePresentUserCallActualArgument(loc, builder, preparedActual, |
1473 | dummyType, arg, callContext); |
1474 | |
1475 | // Conditional dummy argument preparation. The actual may be absent |
1476 | // at runtime, causing any addressing, copy, and packaging to have |
1477 | // undefined behavior. |
1478 | // To simplify the handling of this case, the "normal" dummy preparation |
1479 | // helper is used, except its generated code is wrapped inside a |
1480 | // fir.if(present). |
1481 | mlir::Value isPresent = preparedActual.getIsPresent(); |
1482 | mlir::OpBuilder::InsertPoint insertPt = builder.saveInsertionPoint(); |
1483 | |
1484 | // Code generated in a preparation block that will become the |
1485 | // "then" block in "if (present) then {} else {}". The reason |
1486 | // for this unusual if/then/else generation is that the number |
1487 | // and types of the if results will depend on how the argument |
1488 | // is prepared, and forecasting that here would be brittle. |
1489 | auto badIfOp = builder.create<fir::IfOp>(loc, dummyType, isPresent, |
1490 | /*withElseRegion=*/false); |
1491 | mlir::Block *preparationBlock = &badIfOp.getThenRegion().front(); |
1492 | builder.setInsertionPointToStart(preparationBlock); |
1493 | PreparedDummyArgument unconditionalDummy = |
1494 | preparePresentUserCallActualArgument(loc, builder, preparedActual, |
1495 | dummyType, arg, callContext); |
1496 | builder.restoreInsertionPoint(insertPt); |
1497 | |
1498 | // TODO: when forwarding an optional to an optional of the same kind |
1499 | // (i.e, unconditionalDummy.dummy was not created in preparationBlock), |
1500 | // the if/then/else generation could be skipped to improve the generated |
1501 | // code. |
1502 | |
1503 | // Now that the result types of the ifOp can be deduced, generate |
1504 | // the "real" ifOp (operation result types cannot be changed, so |
1505 | // badIfOp cannot be modified and used here). |
1506 | llvm::SmallVector<mlir::Type> ifOpResultTypes; |
1507 | ConditionallyPreparedDummy conditionalDummy(unconditionalDummy); |
1508 | auto ifOp = builder.create<fir::IfOp>(loc, conditionalDummy.getIfResulTypes(), |
1509 | isPresent, |
1510 | /*withElseRegion=*/true); |
1511 | // Move "preparationBlock" into the "then" of the new |
1512 | // fir.if operation and create fir.result propagating |
1513 | // unconditionalDummy. |
1514 | preparationBlock->moveBefore(&ifOp.getThenRegion().back()); |
1515 | ifOp.getThenRegion().back().erase(); |
1516 | builder.setInsertionPointToEnd(&ifOp.getThenRegion().front()); |
1517 | conditionalDummy.genThenResult(loc, builder); |
1518 | |
1519 | // Generate "else" branch with returning absent values. |
1520 | builder.setInsertionPointToStart(&ifOp.getElseRegion().front()); |
1521 | conditionalDummy.genElseResult(loc, builder); |
1522 | |
1523 | // Build dummy from IfOpResults. |
1524 | builder.setInsertionPointAfter(ifOp); |
1525 | PreparedDummyArgument result = |
1526 | conditionalDummy.getPreparedDummy(ifOp, unconditionalDummy); |
1527 | badIfOp->erase(); |
1528 | return result; |
1529 | } |
1530 | |
1531 | /// Prepare actual argument for a procedure pointer dummy. |
1532 | static PreparedDummyArgument prepareProcedurePointerActualArgument( |
1533 | mlir::Location loc, fir::FirOpBuilder &builder, |
1534 | const Fortran::lower::PreparedActualArgument &preparedActual, |
1535 | mlir::Type dummyType, |
1536 | const Fortran::lower::CallerInterface::PassedEntity &arg, |
1537 | CallContext &callContext) { |
1538 | |
1539 | // NULL() actual to procedure pointer dummy |
1540 | if (Fortran::evaluate::UnwrapExpr<Fortran::evaluate::NullPointer>( |
1541 | *arg.entity) && |
1542 | fir::isBoxProcAddressType(dummyType)) { |
1543 | auto boxTy{Fortran::lower::getUntypedBoxProcType(builder.getContext())}; |
1544 | auto tempBoxProc{builder.createTemporary(loc, boxTy)}; |
1545 | hlfir::Entity nullBoxProc( |
1546 | fir::factory::createNullBoxProc(builder, loc, boxTy)); |
1547 | builder.create<fir::StoreOp>(loc, nullBoxProc, tempBoxProc); |
1548 | return PreparedDummyArgument{tempBoxProc, /*cleanups=*/{}}; |
1549 | } |
1550 | hlfir::Entity actual = preparedActual.getActual(loc, builder); |
1551 | if (actual.isProcedurePointer()) |
1552 | return PreparedDummyArgument{actual, /*cleanups=*/{}}; |
1553 | assert(actual.isProcedure()); |
1554 | // Procedure actual to procedure pointer dummy. |
1555 | auto tempBoxProc{builder.createTemporary(loc, actual.getType())}; |
1556 | builder.create<fir::StoreOp>(loc, actual, tempBoxProc); |
1557 | return PreparedDummyArgument{tempBoxProc, /*cleanups=*/{}}; |
1558 | } |
1559 | |
1560 | /// Prepare arguments of calls to user procedures with actual arguments that |
1561 | /// have been pre-lowered but not yet prepared according to the interface. |
1562 | void prepareUserCallArguments( |
1563 | Fortran::lower::PreparedActualArguments &loweredActuals, |
1564 | Fortran::lower::CallerInterface &caller, mlir::FunctionType callSiteType, |
1565 | CallContext &callContext, llvm::SmallVector<CallCleanUp> &callCleanUps) { |
1566 | using PassBy = Fortran::lower::CallerInterface::PassEntityBy; |
1567 | mlir::Location loc = callContext.loc; |
1568 | bool mustRemapActualToDummyDescriptors = false; |
1569 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
1570 | for (auto [preparedActual, arg] : |
1571 | llvm::zip(loweredActuals, caller.getPassedArguments())) { |
1572 | mlir::Type argTy = callSiteType.getInput(arg.firArgument); |
1573 | if (!preparedActual) { |
1574 | // Optional dummy argument for which there is no actual argument. |
1575 | caller.placeInput(arg, builder.genAbsentOp(loc, argTy)); |
1576 | continue; |
1577 | } |
1578 | |
1579 | switch (arg.passBy) { |
1580 | case PassBy::Value: { |
1581 | // True pass-by-value semantics. |
1582 | assert(!preparedActual->handleDynamicOptional() && "cannot be optional"); |
1583 | hlfir::Entity actual = preparedActual->getActual(loc, builder); |
1584 | hlfir::Entity value = hlfir::loadTrivialScalar(loc, builder, actual); |
1585 | |
1586 | mlir::Type eleTy = value.getFortranElementType(); |
1587 | if (fir::isa_builtin_cptr_type(eleTy)) { |
1588 | // Pass-by-value argument of type(C_PTR/C_FUNPTR). |
1589 | // Load the __address component and pass it by value. |
1590 | if (value.isValue()) { |
1591 | auto associate = hlfir::genAssociateExpr(loc, builder, value, eleTy, |
1592 | "adapt.cptrbyval"); |
1593 | value = hlfir::Entity{genRecordCPtrValueArg( |
1594 | builder, loc, associate.getFirBase(), eleTy)}; |
1595 | builder.create<hlfir::EndAssociateOp>(loc, associate); |
1596 | } else { |
1597 | value = |
1598 | hlfir::Entity{genRecordCPtrValueArg(builder, loc, value, eleTy)}; |
1599 | } |
1600 | } else if (fir::isa_derived(value.getFortranElementType()) || |
1601 | value.isCharacter()) { |
1602 | // BIND(C), VALUE derived type or character. The value must really |
1603 | // be loaded here. |
1604 | auto [exv, cleanup] = hlfir::convertToValue(loc, builder, value); |
1605 | mlir::Value loadedValue = fir::getBase(exv); |
1606 | // Character actual arguments may have unknown length or a length longer |
1607 | // than one. Cast the memory ref to the dummy type so that the load is |
1608 | // valid and only loads what is needed. |
1609 | if (mlir::Type baseTy = fir::dyn_cast_ptrEleTy(loadedValue.getType())) |
1610 | if (fir::isa_char(baseTy)) |
1611 | loadedValue = builder.createConvert( |
1612 | loc, fir::ReferenceType::get(argTy), loadedValue); |
1613 | if (fir::isa_ref_type(loadedValue.getType())) |
1614 | loadedValue = builder.create<fir::LoadOp>(loc, loadedValue); |
1615 | caller.placeInput(arg, loadedValue); |
1616 | if (cleanup) |
1617 | (*cleanup)(); |
1618 | break; |
1619 | } |
1620 | caller.placeInput(arg, builder.createConvert(loc, argTy, value)); |
1621 | } break; |
1622 | case PassBy::BaseAddressValueAttribute: |
1623 | case PassBy::CharBoxValueAttribute: |
1624 | case PassBy::Box: |
1625 | case PassBy::BaseAddress: |
1626 | case PassBy::BoxChar: { |
1627 | PreparedDummyArgument preparedDummy = prepareUserCallActualArgument( |
1628 | loc, builder, *preparedActual, argTy, arg, callContext); |
1629 | callCleanUps.append(preparedDummy.cleanups.rbegin(), |
1630 | preparedDummy.cleanups.rend()); |
1631 | caller.placeInput(arg, preparedDummy.dummy); |
1632 | if (arg.passBy == PassBy::Box) |
1633 | mustRemapActualToDummyDescriptors |= |
1634 | arg.isSequenceAssociatedDescriptor(); |
1635 | } break; |
1636 | case PassBy::BoxProcRef: { |
1637 | PreparedDummyArgument preparedDummy = |
1638 | prepareProcedurePointerActualArgument(loc, builder, *preparedActual, |
1639 | argTy, arg, callContext); |
1640 | callCleanUps.append(preparedDummy.cleanups.rbegin(), |
1641 | preparedDummy.cleanups.rend()); |
1642 | caller.placeInput(arg, preparedDummy.dummy); |
1643 | } break; |
1644 | case PassBy::AddressAndLength: |
1645 | // PassBy::AddressAndLength is only used for character results. Results |
1646 | // are not handled here. |
1647 | fir::emitFatalError( |
1648 | loc, "unexpected PassBy::AddressAndLength for actual arguments"); |
1649 | break; |
1650 | case PassBy::CharProcTuple: { |
1651 | hlfir::Entity actual = preparedActual->getActual(loc, builder); |
1652 | if (actual.isProcedurePointer()) |
1653 | actual = hlfir::derefPointersAndAllocatables(loc, builder, actual); |
1654 | if (!fir::isCharacterProcedureTuple(actual.getType())) |
1655 | actual = fixProcedureDummyMismatch(loc, builder, actual, argTy); |
1656 | caller.placeInput(arg, actual); |
1657 | } break; |
1658 | case PassBy::MutableBox: { |
1659 | const Fortran::lower::SomeExpr *expr = arg.entity->UnwrapExpr(); |
1660 | // C709 and C710. |
1661 | assert(expr && "cannot pass TYPE(*) to POINTER or ALLOCATABLE"); |
1662 | hlfir::Entity actual = preparedActual->getActual(loc, builder); |
1663 | if (Fortran::evaluate::UnwrapExpr<Fortran::evaluate::NullPointer>( |
1664 | *expr)) { |
1665 | // If expr is NULL(), the mutableBox created must be a deallocated |
1666 | // pointer with the dummy argument characteristics (see table 16.5 |
1667 | // in Fortran 2018 standard). |
1668 | // No length parameters are set for the created box because any non |
1669 | // deferred type parameters of the dummy will be evaluated on the |
1670 | // callee side, and it is illegal to use NULL without a MOLD if any |
1671 | // dummy length parameters are assumed. |
1672 | mlir::Type boxTy = fir::dyn_cast_ptrEleTy(argTy); |
1673 | assert(boxTy && mlir::isa<fir::BaseBoxType>(boxTy) && |
1674 | "must be a fir.box type"); |
1675 | mlir::Value boxStorage = |
1676 | fir::factory::genNullBoxStorage(builder, loc, boxTy); |
1677 | caller.placeInput(arg, boxStorage); |
1678 | continue; |
1679 | } |
1680 | if (fir::isPointerType(argTy) && |
1681 | !Fortran::evaluate::IsObjectPointer(*expr)) { |
1682 | // Passing a non POINTER actual argument to a POINTER dummy argument. |
1683 | // Create a pointer of the dummy argument type and assign the actual |
1684 | // argument to it. |
1685 | auto dataTy = llvm::cast<fir::BaseBoxType>(fir::unwrapRefType(argTy)); |
1686 | fir::ExtendedValue actualExv = Fortran::lower::convertToAddress( |
1687 | loc, callContext.converter, actual, callContext.stmtCtx, |
1688 | hlfir::getFortranElementType(dataTy)); |
1689 | // If the dummy is an assumed-rank pointer, allocate a pointer |
1690 | // descriptor with the actual argument rank (if it is not assumed-rank |
1691 | // itself). |
1692 | if (dataTy.isAssumedRank()) { |
1693 | dataTy = |
1694 | dataTy.getBoxTypeWithNewShape(fir::getBase(actualExv).getType()); |
1695 | } |
1696 | mlir::Value irBox = builder.createTemporary(loc, dataTy); |
1697 | fir::MutableBoxValue ptrBox(irBox, |
1698 | /*nonDeferredParams=*/mlir::ValueRange{}, |
1699 | /*mutableProperties=*/{}); |
1700 | fir::factory::associateMutableBox(builder, loc, ptrBox, actualExv, |
1701 | /*lbounds=*/std::nullopt); |
1702 | caller.placeInput(arg, irBox); |
1703 | continue; |
1704 | } |
1705 | // Passing a POINTER to a POINTER, or an ALLOCATABLE to an ALLOCATABLE. |
1706 | assert(actual.isMutableBox() && "actual must be a mutable box"); |
1707 | if (fir::isAllocatableType(argTy) && arg.isIntentOut() && |
1708 | callContext.isBindcCall()) { |
1709 | // INTENT(OUT) allocatables are deallocated on the callee side, |
1710 | // but BIND(C) procedures may be implemented in C, so deallocation is |
1711 | // also done on the caller side (if the procedure is implemented in |
1712 | // Fortran, the deallocation attempt in the callee will be a no-op). |
1713 | auto [exv, cleanup] = |
1714 | hlfir::translateToExtendedValue(loc, builder, actual); |
1715 | const auto *mutableBox = exv.getBoxOf<fir::MutableBoxValue>(); |
1716 | assert(mutableBox && !cleanup && "expect allocatable"); |
1717 | Fortran::lower::genDeallocateIfAllocated(callContext.converter, |
1718 | *mutableBox, loc); |
1719 | } |
1720 | caller.placeInput(arg, actual); |
1721 | } break; |
1722 | } |
1723 | } |
1724 | |
1725 | // Handle cases where caller must allocate the result or a fir.box for it. |
1726 | if (mustRemapActualToDummyDescriptors) |
1727 | remapActualToDummyDescriptors(loc, callContext.converter, |
1728 | callContext.symMap, loweredActuals, caller, |
1729 | callContext.isBindcCall()); |
1730 | } |
1731 | |
1732 | /// Lower calls to user procedures with actual arguments that have been |
1733 | /// pre-lowered but not yet prepared according to the interface. |
1734 | /// This can be called for elemental procedures, but only with scalar |
1735 | /// arguments: if there are array arguments, it must be provided with |
1736 | /// the array argument elements value and will return the corresponding |
1737 | /// scalar result value. |
1738 | static std::optional<hlfir::EntityWithAttributes> |
1739 | genUserCall(Fortran::lower::PreparedActualArguments &loweredActuals, |
1740 | Fortran::lower::CallerInterface &caller, |
1741 | mlir::FunctionType callSiteType, CallContext &callContext) { |
1742 | mlir::Location loc = callContext.loc; |
1743 | llvm::SmallVector<CallCleanUp> callCleanUps; |
1744 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
1745 | |
1746 | prepareUserCallArguments(loweredActuals, caller, callSiteType, callContext, |
1747 | callCleanUps); |
1748 | |
1749 | // Prepare lowered arguments according to the interface |
1750 | // and map the lowered values to the dummy |
1751 | // arguments. |
1752 | auto [loweredResult, resultIsFinalized] = Fortran::lower::genCallOpAndResult( |
1753 | loc, callContext.converter, callContext.symMap, callContext.stmtCtx, |
1754 | caller, callSiteType, callContext.resultType, |
1755 | callContext.isElementalProcWithArrayArgs()); |
1756 | |
1757 | // Clean-up associations and copy-in. |
1758 | // The association clean-ups are postponed to the end of the statement |
1759 | // lowering. The copy-in clean-ups may be delayed as well, |
1760 | // but they are done immediately after the call currently. |
1761 | llvm::SmallVector<CallCleanUp> associateCleanups; |
1762 | for (auto cleanUp : callCleanUps) { |
1763 | auto postponed = |
1764 | cleanUp.genCleanUp(loc, builder, /*postponeAssociates=*/true); |
1765 | if (postponed) |
1766 | associateCleanups.push_back(Elt: *postponed); |
1767 | } |
1768 | |
1769 | fir::FirOpBuilder *bldr = &builder; |
1770 | callContext.stmtCtx.attachCleanup([=]() { |
1771 | for (auto cleanUp : associateCleanups) |
1772 | (void)cleanUp.genCleanUp(loc, *bldr, /*postponeAssociates=*/false); |
1773 | }); |
1774 | if (auto *entity = std::get_if<hlfir::EntityWithAttributes>(&loweredResult)) |
1775 | return *entity; |
1776 | |
1777 | auto &result = std::get<fir::ExtendedValue>(loweredResult); |
1778 | |
1779 | // For procedure pointer function result, just return the call. |
1780 | if (callContext.resultType && |
1781 | mlir::isa<fir::BoxProcType>(*callContext.resultType)) |
1782 | return hlfir::EntityWithAttributes(fir::getBase(result)); |
1783 | |
1784 | if (!fir::getBase(result)) |
1785 | return std::nullopt; // subroutine call. |
1786 | |
1787 | if (fir::isPointerType(fir::getBase(result).getType())) |
1788 | return extendedValueToHlfirEntity(loc, builder, result, tempResultName); |
1789 | |
1790 | if (!resultIsFinalized) { |
1791 | hlfir::Entity resultEntity = |
1792 | extendedValueToHlfirEntity(loc, builder, result, tempResultName); |
1793 | resultEntity = loadTrivialScalar(loc, builder, resultEntity); |
1794 | if (resultEntity.isVariable()) { |
1795 | // If the result has no finalization, it can be moved into an expression. |
1796 | // In such case, the expression should not be freed after its use since |
1797 | // the result is stack allocated or deallocation (for allocatable results) |
1798 | // was already inserted in genCallOpAndResult. |
1799 | auto asExpr = builder.create<hlfir::AsExprOp>( |
1800 | loc, resultEntity, /*mustFree=*/builder.createBool(loc, false)); |
1801 | return hlfir::EntityWithAttributes{asExpr.getResult()}; |
1802 | } |
1803 | return hlfir::EntityWithAttributes{resultEntity}; |
1804 | } |
1805 | // If the result has finalization, it cannot be moved because use of its |
1806 | // value have been created in the statement context and may be emitted |
1807 | // after the hlfir.expr destroy, so the result is kept as a variable in |
1808 | // HLFIR. This may lead to copies when passing the result to an argument |
1809 | // with VALUE, and this do not convey the fact that the result will not |
1810 | // change, but is correct, and using hlfir.expr without the move would |
1811 | // trigger a copy that may be avoided. |
1812 | |
1813 | // Load allocatable results before emitting the hlfir.declare and drop its |
1814 | // lower bounds: this is not a variable From the Fortran point of view, so |
1815 | // the lower bounds are ones when inquired on the caller side. |
1816 | const auto *allocatable = result.getBoxOf<fir::MutableBoxValue>(); |
1817 | fir::ExtendedValue loadedResult = |
1818 | allocatable |
1819 | ? fir::factory::genMutableBoxRead(builder, loc, *allocatable, |
1820 | /*mayBePolymorphic=*/true, |
1821 | /*preserveLowerBounds=*/false) |
1822 | : result; |
1823 | return extendedValueToHlfirEntity(loc, builder, loadedResult, tempResultName); |
1824 | } |
1825 | |
1826 | /// Create an optional dummy argument value from an entity that may be |
1827 | /// absent. \p actualGetter callback returns hlfir::Entity denoting |
1828 | /// the lowered actual argument. \p actualGetter can only return numerical |
1829 | /// or logical scalar entity. |
1830 | /// If the entity is considered absent according to 15.5.2.12 point 1., the |
1831 | /// returned value is zero (or false), otherwise it is the value of the entity. |
1832 | /// \p eleType specifies the entity's Fortran element type. |
1833 | template <typename T> |
1834 | static ExvAndCleanup genOptionalValue(fir::FirOpBuilder &builder, |
1835 | mlir::Location loc, mlir::Type eleType, |
1836 | T actualGetter, mlir::Value isPresent) { |
1837 | return {builder |
1838 | .genIfOp(loc, {eleType}, isPresent, |
1839 | /*withElseRegion=*/true) |
1840 | .genThen([&]() { |
1841 | hlfir::Entity entity = actualGetter(loc, builder); |
1842 | assert(eleType == entity.getFortranElementType() && |
1843 | "result type mismatch in genOptionalValue"); |
1844 | assert(entity.isScalar() && fir::isa_trivial(eleType) && |
1845 | "must be a numerical or logical scalar"); |
1846 | mlir::Value val = |
1847 | hlfir::loadTrivialScalar(loc, builder, entity); |
1848 | builder.create<fir::ResultOp>(loc, val); |
1849 | }) |
1850 | .genElse([&]() { |
1851 | mlir::Value zero = |
1852 | fir::factory::createZeroValue(builder, loc, eleType); |
1853 | builder.create<fir::ResultOp>(loc, zero); |
1854 | }) |
1855 | .getResults()[0], |
1856 | std::nullopt}; |
1857 | } |
1858 | |
1859 | /// Create an optional dummy argument address from \p entity that may be |
1860 | /// absent. If \p entity is considered absent according to 15.5.2.12 point 1., |
1861 | /// the returned value is a null pointer, otherwise it is the address of \p |
1862 | /// entity. |
1863 | static ExvAndCleanup genOptionalAddr(fir::FirOpBuilder &builder, |
1864 | mlir::Location loc, hlfir::Entity entity, |
1865 | mlir::Value isPresent) { |
1866 | auto [exv, cleanup] = hlfir::translateToExtendedValue(loc, builder, entity); |
1867 | // If it is an exv pointer/allocatable, then it cannot be absent |
1868 | // because it is passed to a non-pointer/non-allocatable. |
1869 | if (const auto *box = exv.getBoxOf<fir::MutableBoxValue>()) |
1870 | return {fir::factory::genMutableBoxRead(builder, loc, *box), cleanup}; |
1871 | // If this is not a POINTER or ALLOCATABLE, then it is already an OPTIONAL |
1872 | // address and can be passed directly. |
1873 | return {exv, cleanup}; |
1874 | } |
1875 | |
1876 | /// Create an optional dummy argument address from \p entity that may be |
1877 | /// absent. If \p entity is considered absent according to 15.5.2.12 point 1., |
1878 | /// the returned value is an absent fir.box, otherwise it is a fir.box |
1879 | /// describing \p entity. |
1880 | static ExvAndCleanup genOptionalBox(fir::FirOpBuilder &builder, |
1881 | mlir::Location loc, hlfir::Entity entity, |
1882 | mlir::Value isPresent) { |
1883 | auto [exv, cleanup] = hlfir::translateToExtendedValue(loc, builder, entity); |
1884 | |
1885 | // Non allocatable/pointer optional box -> simply forward |
1886 | if (exv.getBoxOf<fir::BoxValue>()) |
1887 | return {exv, cleanup}; |
1888 | |
1889 | fir::ExtendedValue newExv = exv; |
1890 | // Optional allocatable/pointer -> Cannot be absent, but need to translate |
1891 | // unallocated/diassociated into absent fir.box. |
1892 | if (const auto *box = exv.getBoxOf<fir::MutableBoxValue>()) |
1893 | newExv = fir::factory::genMutableBoxRead(builder, loc, *box); |
1894 | |
1895 | // createBox will not do create any invalid memory dereferences if exv is |
1896 | // absent. The created fir.box will not be usable, but the SelectOp below |
1897 | // ensures it won't be. |
1898 | mlir::Value box = builder.createBox(loc, newExv); |
1899 | mlir::Type boxType = box.getType(); |
1900 | auto absent = builder.create<fir::AbsentOp>(loc, boxType); |
1901 | auto boxOrAbsent = builder.create<mlir::arith::SelectOp>( |
1902 | loc, boxType, isPresent, box, absent); |
1903 | return {fir::BoxValue(boxOrAbsent), cleanup}; |
1904 | } |
1905 | |
1906 | /// Lower calls to intrinsic procedures with custom optional handling where the |
1907 | /// actual arguments have been pre-lowered |
1908 | static std::optional<hlfir::EntityWithAttributes> genCustomIntrinsicRefCore( |
1909 | Fortran::lower::PreparedActualArguments &loweredActuals, |
1910 | const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
1911 | CallContext &callContext) { |
1912 | auto &builder = callContext.getBuilder(); |
1913 | const auto &loc = callContext.loc; |
1914 | assert(intrinsic && |
1915 | Fortran::lower::intrinsicRequiresCustomOptionalHandling( |
1916 | callContext.procRef, *intrinsic, callContext.converter)); |
1917 | |
1918 | // helper to get a particular prepared argument |
1919 | auto getArgument = [&](std::size_t i, bool loadArg) -> fir::ExtendedValue { |
1920 | if (!loweredActuals[i]) |
1921 | return fir::getAbsentIntrinsicArgument(); |
1922 | hlfir::Entity actual = loweredActuals[i]->getActual(loc, builder); |
1923 | if (loadArg && fir::conformsWithPassByRef(actual.getType())) { |
1924 | return hlfir::loadTrivialScalar(loc, builder, actual); |
1925 | } |
1926 | return Fortran::lower::translateToExtendedValue(loc, builder, actual, |
1927 | callContext.stmtCtx); |
1928 | }; |
1929 | // helper to get the isPresent flag for a particular prepared argument |
1930 | auto isPresent = [&](std::size_t i) -> std::optional<mlir::Value> { |
1931 | if (!loweredActuals[i]) |
1932 | return {builder.createBool(loc, false)}; |
1933 | if (loweredActuals[i]->handleDynamicOptional()) |
1934 | return {loweredActuals[i]->getIsPresent()}; |
1935 | return std::nullopt; |
1936 | }; |
1937 | |
1938 | assert(callContext.resultType && |
1939 | "the elemental intrinsics with custom handling are all functions"); |
1940 | // if callContext.resultType is an array then this was originally an elemental |
1941 | // call. What we are lowering here is inside the kernel of the hlfir.elemental |
1942 | // so we should return the scalar type. If the return type is already a scalar |
1943 | // then it should be unchanged here. |
1944 | mlir::Type resTy = hlfir::getFortranElementType(*callContext.resultType); |
1945 | fir::ExtendedValue result = Fortran::lower::lowerCustomIntrinsic( |
1946 | builder, loc, callContext.getProcedureName(), resTy, isPresent, |
1947 | getArgument, loweredActuals.size(), callContext.stmtCtx); |
1948 | |
1949 | return {hlfir::EntityWithAttributes{extendedValueToHlfirEntity( |
1950 | loc, builder, result, ".tmp.custom_intrinsic_result")}}; |
1951 | } |
1952 | |
1953 | /// Lower calls to intrinsic procedures with actual arguments that have been |
1954 | /// pre-lowered but have not yet been prepared according to the interface. |
1955 | static std::optional<hlfir::EntityWithAttributes> |
1956 | genIntrinsicRefCore(Fortran::lower::PreparedActualArguments &loweredActuals, |
1957 | const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
1958 | const fir::IntrinsicHandlerEntry &intrinsicEntry, |
1959 | CallContext &callContext) { |
1960 | auto &converter = callContext.converter; |
1961 | if (intrinsic && Fortran::lower::intrinsicRequiresCustomOptionalHandling( |
1962 | callContext.procRef, *intrinsic, converter)) |
1963 | return genCustomIntrinsicRefCore(loweredActuals, intrinsic, callContext); |
1964 | llvm::SmallVector<fir::ExtendedValue> operands; |
1965 | llvm::SmallVector<hlfir::CleanupFunction> cleanupFns; |
1966 | auto addToCleanups = [&cleanupFns](std::optional<hlfir::CleanupFunction> fn) { |
1967 | if (fn) |
1968 | cleanupFns.emplace_back(std::move(*fn)); |
1969 | }; |
1970 | auto &stmtCtx = callContext.stmtCtx; |
1971 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
1972 | mlir::Location loc = callContext.loc; |
1973 | const fir::IntrinsicArgumentLoweringRules *argLowering = |
1974 | intrinsicEntry.getArgumentLoweringRules(); |
1975 | for (auto arg : llvm::enumerate(loweredActuals)) { |
1976 | if (!arg.value()) { |
1977 | operands.emplace_back(fir::getAbsentIntrinsicArgument()); |
1978 | continue; |
1979 | } |
1980 | if (!argLowering) { |
1981 | // No argument lowering instruction, lower by value. |
1982 | assert(!arg.value()->handleDynamicOptional() && |
1983 | "should use genOptionalValue"); |
1984 | hlfir::Entity actual = arg.value()->getActual(loc, builder); |
1985 | operands.emplace_back( |
1986 | Fortran::lower::convertToValue(loc, converter, actual, stmtCtx)); |
1987 | continue; |
1988 | } |
1989 | // Helper to get the type of the Fortran expression in case it is a |
1990 | // computed value that must be placed in memory (logicals are computed as |
1991 | // i1, but must be placed in memory as fir.logical). |
1992 | auto getActualFortranElementType = [&]() -> mlir::Type { |
1993 | if (const Fortran::lower::SomeExpr *expr = |
1994 | callContext.procRef.UnwrapArgExpr(arg.index())) { |
1995 | |
1996 | mlir::Type type = converter.genType(*expr); |
1997 | return hlfir::getFortranElementType(type); |
1998 | } |
1999 | // TYPE(*): is already in memory anyway. Can return none |
2000 | // here. |
2001 | return builder.getNoneType(); |
2002 | }; |
2003 | // Ad-hoc argument lowering handling. |
2004 | fir::ArgLoweringRule argRules = |
2005 | fir::lowerIntrinsicArgumentAs(*argLowering, arg.index()); |
2006 | if (arg.value()->handleDynamicOptional()) { |
2007 | mlir::Value isPresent = arg.value()->getIsPresent(); |
2008 | switch (argRules.lowerAs) { |
2009 | case fir::LowerIntrinsicArgAs::Value: { |
2010 | // In case of elemental call, getActual() may produce |
2011 | // a designator denoting the array element to be passed |
2012 | // to the subprogram. If the actual array is dynamically |
2013 | // optional the designator must be generated under |
2014 | // isPresent check, because the box bounds reads will be |
2015 | // generated in the codegen. These reads are illegal, |
2016 | // if the dynamically optional argument is absent. |
2017 | auto getActualCb = [&](mlir::Location loc, |
2018 | fir::FirOpBuilder &builder) -> hlfir::Entity { |
2019 | return arg.value()->getActual(loc, builder); |
2020 | }; |
2021 | auto [exv, cleanup] = |
2022 | genOptionalValue(builder, loc, getActualFortranElementType(), |
2023 | getActualCb, isPresent); |
2024 | addToCleanups(std::move(cleanup)); |
2025 | operands.emplace_back(exv); |
2026 | continue; |
2027 | } |
2028 | case fir::LowerIntrinsicArgAs::Addr: { |
2029 | hlfir::Entity actual = arg.value()->getActual(loc, builder); |
2030 | auto [exv, cleanup] = genOptionalAddr(builder, loc, actual, isPresent); |
2031 | addToCleanups(std::move(cleanup)); |
2032 | operands.emplace_back(exv); |
2033 | continue; |
2034 | } |
2035 | case fir::LowerIntrinsicArgAs::Box: { |
2036 | hlfir::Entity actual = arg.value()->getActual(loc, builder); |
2037 | auto [exv, cleanup] = genOptionalBox(builder, loc, actual, isPresent); |
2038 | addToCleanups(std::move(cleanup)); |
2039 | operands.emplace_back(exv); |
2040 | continue; |
2041 | } |
2042 | case fir::LowerIntrinsicArgAs::Inquired: { |
2043 | hlfir::Entity actual = arg.value()->getActual(loc, builder); |
2044 | auto [exv, cleanup] = |
2045 | hlfir::translateToExtendedValue(loc, builder, actual); |
2046 | addToCleanups(std::move(cleanup)); |
2047 | operands.emplace_back(exv); |
2048 | continue; |
2049 | } |
2050 | } |
2051 | llvm_unreachable("bad switch"); |
2052 | } |
2053 | |
2054 | hlfir::Entity actual = arg.value()->getActual(loc, builder); |
2055 | switch (argRules.lowerAs) { |
2056 | case fir::LowerIntrinsicArgAs::Value: |
2057 | operands.emplace_back( |
2058 | Fortran::lower::convertToValue(loc, converter, actual, stmtCtx)); |
2059 | continue; |
2060 | case fir::LowerIntrinsicArgAs::Addr: |
2061 | operands.emplace_back(Fortran::lower::convertToAddress( |
2062 | loc, converter, actual, stmtCtx, getActualFortranElementType())); |
2063 | continue; |
2064 | case fir::LowerIntrinsicArgAs::Box: |
2065 | operands.emplace_back(Fortran::lower::convertToBox( |
2066 | loc, converter, actual, stmtCtx, getActualFortranElementType())); |
2067 | continue; |
2068 | case fir::LowerIntrinsicArgAs::Inquired: |
2069 | if (const Fortran::lower::SomeExpr *expr = |
2070 | callContext.procRef.UnwrapArgExpr(arg.index())) { |
2071 | if (Fortran::evaluate::UnwrapExpr<Fortran::evaluate::NullPointer>( |
2072 | *expr)) { |
2073 | // NULL() pointer without a MOLD must be passed as a deallocated |
2074 | // pointer (see table 16.5 in Fortran 2018 standard). |
2075 | // !fir.box<!fir.ptr<none>> should always be valid in this context. |
2076 | mlir::Type noneTy = mlir::NoneType::get(builder.getContext()); |
2077 | mlir::Type nullPtrTy = fir::PointerType::get(noneTy); |
2078 | mlir::Type boxTy = fir::BoxType::get(nullPtrTy); |
2079 | mlir::Value boxStorage = |
2080 | fir::factory::genNullBoxStorage(builder, loc, boxTy); |
2081 | hlfir::EntityWithAttributes nullBoxEntity = |
2082 | extendedValueToHlfirEntity(loc, builder, boxStorage, |
2083 | ".tmp.null_box"); |
2084 | operands.emplace_back(Fortran::lower::translateToExtendedValue( |
2085 | loc, builder, nullBoxEntity, stmtCtx)); |
2086 | continue; |
2087 | } |
2088 | } |
2089 | // Place hlfir.expr in memory, and unbox fir.boxchar. Other entities |
2090 | // are translated to fir::ExtendedValue without transformation (notably, |
2091 | // pointers/allocatable are not dereferenced). |
2092 | // TODO: once lowering to FIR retires, UBOUND and LBOUND can be simplified |
2093 | // since the fir.box lowered here are now guaranteed to contain the local |
2094 | // lower bounds thanks to the hlfir.declare (the extra rebox can be |
2095 | // removed). |
2096 | operands.emplace_back(Fortran::lower::translateToExtendedValue( |
2097 | loc, builder, actual, stmtCtx)); |
2098 | continue; |
2099 | } |
2100 | llvm_unreachable("bad switch"); |
2101 | } |
2102 | // genIntrinsicCall needs the scalar type, even if this is a transformational |
2103 | // procedure returning an array. |
2104 | std::optional<mlir::Type> scalarResultType; |
2105 | if (callContext.resultType) |
2106 | scalarResultType = hlfir::getFortranElementType(*callContext.resultType); |
2107 | const std::string intrinsicName = callContext.getProcedureName(); |
2108 | // Let the intrinsic library lower the intrinsic procedure call. |
2109 | auto [resultExv, mustBeFreed] = genIntrinsicCall( |
2110 | builder, loc, intrinsicEntry, scalarResultType, operands, &converter); |
2111 | for (const hlfir::CleanupFunction &fn : cleanupFns) |
2112 | fn(); |
2113 | if (!fir::getBase(resultExv)) |
2114 | return std::nullopt; |
2115 | hlfir::EntityWithAttributes resultEntity = extendedValueToHlfirEntity( |
2116 | loc, builder, resultExv, ".tmp.intrinsic_result"); |
2117 | // Move result into memory into an hlfir.expr since they are immutable from |
2118 | // that point, and the result storage is some temp. "Null" is special: it |
2119 | // returns a null pointer variable that should not be transformed into a value |
2120 | // (what matters is the memory address). |
2121 | if (resultEntity.isVariable() && intrinsicName != "null") { |
2122 | assert(!fir::isa_trivial(fir::unwrapRefType(resultEntity.getType())) && |
2123 | "expect intrinsic scalar results to not be in memory"); |
2124 | hlfir::AsExprOp asExpr; |
2125 | // Character/Derived MERGE lowering returns one of its argument address |
2126 | // (this is the only intrinsic implemented in that way so far). The |
2127 | // ownership of this address cannot be taken here since it may not be a |
2128 | // temp. |
2129 | if (intrinsicName == "merge") |
2130 | asExpr = builder.create<hlfir::AsExprOp>(loc, resultEntity); |
2131 | else |
2132 | asExpr = builder.create<hlfir::AsExprOp>( |
2133 | loc, resultEntity, builder.createBool(loc, mustBeFreed)); |
2134 | resultEntity = hlfir::EntityWithAttributes{asExpr.getResult()}; |
2135 | } |
2136 | return resultEntity; |
2137 | } |
2138 | |
2139 | /// Lower calls to intrinsic procedures with actual arguments that have been |
2140 | /// pre-lowered but have not yet been prepared according to the interface. |
2141 | static std::optional<hlfir::EntityWithAttributes> genHLFIRIntrinsicRefCore( |
2142 | Fortran::lower::PreparedActualArguments &loweredActuals, |
2143 | const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2144 | const fir::IntrinsicHandlerEntry &intrinsicEntry, |
2145 | CallContext &callContext) { |
2146 | // Try lowering transformational intrinsic ops to HLFIR ops if enabled |
2147 | // (transformational always have a result type) |
2148 | if (useHlfirIntrinsicOps && callContext.resultType) { |
2149 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2150 | mlir::Location loc = callContext.loc; |
2151 | const std::string intrinsicName = callContext.getProcedureName(); |
2152 | const fir::IntrinsicArgumentLoweringRules *argLowering = |
2153 | intrinsicEntry.getArgumentLoweringRules(); |
2154 | std::optional<hlfir::EntityWithAttributes> res = |
2155 | Fortran::lower::lowerHlfirIntrinsic(builder, loc, intrinsicName, |
2156 | loweredActuals, argLowering, |
2157 | *callContext.resultType); |
2158 | if (res) |
2159 | return res; |
2160 | } |
2161 | |
2162 | // fallback to calling the intrinsic via fir.call |
2163 | return genIntrinsicRefCore(loweredActuals, intrinsic, intrinsicEntry, |
2164 | callContext); |
2165 | } |
2166 | |
2167 | namespace { |
2168 | template <typename ElementalCallBuilderImpl> |
2169 | class ElementalCallBuilder { |
2170 | public: |
2171 | std::optional<hlfir::EntityWithAttributes> |
2172 | genElementalCall(Fortran::lower::PreparedActualArguments &loweredActuals, |
2173 | bool isImpure, CallContext &callContext) { |
2174 | mlir::Location loc = callContext.loc; |
2175 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2176 | unsigned numArgs = loweredActuals.size(); |
2177 | // Step 1: dereference pointers/allocatables and compute elemental shape. |
2178 | mlir::Value shape; |
2179 | Fortran::lower::PreparedActualArgument *optionalWithShape; |
2180 | // 10.1.4 p5. Impure elemental procedures must be called in element order. |
2181 | bool mustBeOrdered = isImpure; |
2182 | for (unsigned i = 0; i < numArgs; ++i) { |
2183 | auto &preparedActual = loweredActuals[i]; |
2184 | if (preparedActual) { |
2185 | // Elemental procedure dummy arguments cannot be pointer/allocatables |
2186 | // (C15100), so it is safe to dereference any pointer or allocatable |
2187 | // actual argument now instead of doing this inside the elemental |
2188 | // region. |
2189 | preparedActual->derefPointersAndAllocatables(loc, builder); |
2190 | // Better to load scalars outside of the loop when possible. |
2191 | if (!preparedActual->handleDynamicOptional() && |
2192 | impl().canLoadActualArgumentBeforeLoop(i)) |
2193 | preparedActual->loadTrivialScalar(loc, builder); |
2194 | // TODO: merge shape instead of using the first one. |
2195 | if (!shape && preparedActual->isArray()) { |
2196 | if (preparedActual->handleDynamicOptional()) |
2197 | optionalWithShape = &*preparedActual; |
2198 | else |
2199 | shape = preparedActual->genShape(loc, builder); |
2200 | } |
2201 | // 15.8.3 p1. Elemental procedure with intent(out)/intent(inout) |
2202 | // arguments must be called in element order. |
2203 | if (impl().argMayBeModifiedByCall(i)) |
2204 | mustBeOrdered = true; |
2205 | } |
2206 | } |
2207 | if (!shape && optionalWithShape) { |
2208 | // If all array operands appear in optional positions, then none of them |
2209 | // is allowed to be absent as per 15.5.2.12 point 3. (6). Just pick the |
2210 | // first operand. |
2211 | shape = optionalWithShape->genShape(loc, builder); |
2212 | // TODO: There is an opportunity to add a runtime check here that |
2213 | // this array is present as required. Also, the optionality of all actual |
2214 | // could be checked and reset given the Fortran requirement. |
2215 | optionalWithShape->resetOptionalAspect(); |
2216 | } |
2217 | assert(shape && |
2218 | "elemental array calls must have at least one array arguments"); |
2219 | |
2220 | // Evaluate the actual argument array expressions before the elemental |
2221 | // call of an impure subprogram or a subprogram with intent(out) or |
2222 | // intent(inout) arguments. Note that the scalar arguments are handled |
2223 | // above. |
2224 | if (mustBeOrdered) { |
2225 | for (auto &preparedActual : loweredActuals) { |
2226 | if (preparedActual) { |
2227 | if (hlfir::AssociateOp associate = |
2228 | preparedActual->associateIfArrayExpr(loc, builder)) { |
2229 | fir::FirOpBuilder *bldr = &builder; |
2230 | callContext.stmtCtx.attachCleanup( |
2231 | [=]() { bldr->create<hlfir::EndAssociateOp>(loc, associate); }); |
2232 | } |
2233 | } |
2234 | } |
2235 | } |
2236 | |
2237 | // Push a new local scope so that any temps made inside the elemental |
2238 | // iterations are cleaned up inside the iterations. |
2239 | if (!callContext.resultType) { |
2240 | // Subroutine case. Generate call inside loop nest. |
2241 | hlfir::LoopNest loopNest = |
2242 | hlfir::genLoopNest(loc, builder, shape, !mustBeOrdered); |
2243 | mlir::ValueRange oneBasedIndices = loopNest.oneBasedIndices; |
2244 | auto insPt = builder.saveInsertionPoint(); |
2245 | builder.setInsertionPointToStart(loopNest.body); |
2246 | callContext.stmtCtx.pushScope(); |
2247 | for (auto &preparedActual : loweredActuals) |
2248 | if (preparedActual) |
2249 | preparedActual->setElementalIndices(oneBasedIndices); |
2250 | impl().genElementalKernel(loweredActuals, callContext); |
2251 | callContext.stmtCtx.finalizeAndPop(); |
2252 | builder.restoreInsertionPoint(insPt); |
2253 | return std::nullopt; |
2254 | } |
2255 | // Function case: generate call inside hlfir.elemental |
2256 | mlir::Type elementType = |
2257 | hlfir::getFortranElementType(*callContext.resultType); |
2258 | // Get result length parameters. |
2259 | llvm::SmallVector<mlir::Value> typeParams; |
2260 | if (mlir::isa<fir::CharacterType>(elementType) || |
2261 | fir::isRecordWithTypeParameters(elementType)) { |
2262 | auto charType = mlir::dyn_cast<fir::CharacterType>(elementType); |
2263 | if (charType && charType.hasConstantLen()) |
2264 | typeParams.push_back(builder.createIntegerConstant( |
2265 | loc, builder.getIndexType(), charType.getLen())); |
2266 | else if (charType) |
2267 | typeParams.push_back(impl().computeDynamicCharacterResultLength( |
2268 | loweredActuals, callContext)); |
2269 | else |
2270 | TODO( |
2271 | loc, |
2272 | "compute elemental PDT function result length parameters in HLFIR"); |
2273 | } |
2274 | auto genKernel = [&](mlir::Location l, fir::FirOpBuilder &b, |
2275 | mlir::ValueRange oneBasedIndices) -> hlfir::Entity { |
2276 | callContext.stmtCtx.pushScope(); |
2277 | for (auto &preparedActual : loweredActuals) |
2278 | if (preparedActual) |
2279 | preparedActual->setElementalIndices(oneBasedIndices); |
2280 | auto res = *impl().genElementalKernel(loweredActuals, callContext); |
2281 | callContext.stmtCtx.finalizeAndPop(); |
2282 | // Note that an hlfir.destroy is not emitted for the result since it |
2283 | // is still used by the hlfir.yield_element that also marks its last |
2284 | // use. |
2285 | return res; |
2286 | }; |
2287 | mlir::Value polymorphicMold; |
2288 | if (fir::isPolymorphicType(*callContext.resultType)) |
2289 | polymorphicMold = |
2290 | impl().getPolymorphicResultMold(loweredActuals, callContext); |
2291 | mlir::Value elemental = |
2292 | hlfir::genElementalOp(loc, builder, elementType, shape, typeParams, |
2293 | genKernel, !mustBeOrdered, polymorphicMold); |
2294 | // If the function result requires finalization, then it has to be done |
2295 | // for the array result of the elemental call. We have to communicate |
2296 | // this via the DestroyOp's attribute. |
2297 | bool mustFinalizeExpr = impl().resultMayRequireFinalization(callContext); |
2298 | fir::FirOpBuilder *bldr = &builder; |
2299 | callContext.stmtCtx.attachCleanup([=]() { |
2300 | bldr->create<hlfir::DestroyOp>(loc, elemental, mustFinalizeExpr); |
2301 | }); |
2302 | return hlfir::EntityWithAttributes{elemental}; |
2303 | } |
2304 | |
2305 | private: |
2306 | ElementalCallBuilderImpl &impl() { |
2307 | return *static_cast<ElementalCallBuilderImpl *>(this); |
2308 | } |
2309 | }; |
2310 | |
2311 | class ElementalUserCallBuilder |
2312 | : public ElementalCallBuilder<ElementalUserCallBuilder> { |
2313 | public: |
2314 | ElementalUserCallBuilder(Fortran::lower::CallerInterface &caller, |
2315 | mlir::FunctionType callSiteType) |
2316 | : caller{caller}, callSiteType{callSiteType} {} |
2317 | std::optional<hlfir::Entity> |
2318 | genElementalKernel(Fortran::lower::PreparedActualArguments &loweredActuals, |
2319 | CallContext &callContext) { |
2320 | return genUserCall(loweredActuals, caller, callSiteType, callContext); |
2321 | } |
2322 | |
2323 | bool argMayBeModifiedByCall(unsigned argIdx) const { |
2324 | assert(argIdx < caller.getPassedArguments().size() && "bad argument index"); |
2325 | return caller.getPassedArguments()[argIdx].mayBeModifiedByCall(); |
2326 | } |
2327 | |
2328 | bool canLoadActualArgumentBeforeLoop(unsigned argIdx) const { |
2329 | using PassBy = Fortran::lower::CallerInterface::PassEntityBy; |
2330 | const auto &passedArgs{caller.getPassedArguments()}; |
2331 | assert(argIdx < passedArgs.size() && "bad argument index"); |
2332 | // If the actual argument does not need to be passed via an address, |
2333 | // or will be passed in the address of a temporary copy, it can be loaded |
2334 | // before the elemental loop nest. |
2335 | const auto &arg{passedArgs[argIdx]}; |
2336 | return arg.passBy == PassBy::Value || |
2337 | arg.passBy == PassBy::BaseAddressValueAttribute; |
2338 | } |
2339 | |
2340 | mlir::Value computeDynamicCharacterResultLength( |
2341 | Fortran::lower::PreparedActualArguments &loweredActuals, |
2342 | CallContext &callContext) { |
2343 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2344 | mlir::Location loc = callContext.loc; |
2345 | auto &converter = callContext.converter; |
2346 | mlir::Type idxTy = builder.getIndexType(); |
2347 | llvm::SmallVector<CallCleanUp> callCleanUps; |
2348 | |
2349 | prepareUserCallArguments(loweredActuals, caller, callSiteType, callContext, |
2350 | callCleanUps); |
2351 | |
2352 | callContext.symMap.pushScope(); |
2353 | |
2354 | // Map prepared argument to dummy symbol to be able to lower spec expr. |
2355 | for (const auto &arg : caller.getPassedArguments()) { |
2356 | const Fortran::semantics::Symbol *sym = caller.getDummySymbol(arg); |
2357 | assert(sym && "expect symbol for dummy argument"); |
2358 | auto input = caller.getInput(arg); |
2359 | fir::ExtendedValue exv = Fortran::lower::translateToExtendedValue( |
2360 | loc, builder, hlfir::Entity{input}, callContext.stmtCtx); |
2361 | fir::FortranVariableOpInterface variableIface = hlfir::genDeclare( |
2362 | loc, builder, exv, "dummy.tmp", fir::FortranVariableFlagsAttr{}); |
2363 | callContext.symMap.addVariableDefinition(*sym, variableIface); |
2364 | } |
2365 | |
2366 | auto lowerSpecExpr = [&](const auto &expr) -> mlir::Value { |
2367 | mlir::Value convertExpr = builder.createConvert( |
2368 | loc, idxTy, |
2369 | fir::getBase(converter.genExprValue(expr, callContext.stmtCtx))); |
2370 | return fir::factory::genMaxWithZero(builder, loc, convertExpr); |
2371 | }; |
2372 | |
2373 | llvm::SmallVector<mlir::Value> lengths; |
2374 | caller.walkResultLengths( |
2375 | [&](const Fortran::lower::SomeExpr &e, bool isAssumedSizeExtent) { |
2376 | assert(!isAssumedSizeExtent && "result cannot be assumed-size"); |
2377 | lengths.emplace_back(lowerSpecExpr(e)); |
2378 | }); |
2379 | callContext.symMap.popScope(); |
2380 | assert(lengths.size() == 1 && "expect 1 length parameter for the result"); |
2381 | return lengths[0]; |
2382 | } |
2383 | |
2384 | mlir::Value getPolymorphicResultMold( |
2385 | Fortran::lower::PreparedActualArguments &loweredActuals, |
2386 | CallContext &callContext) { |
2387 | fir::emitFatalError(callContext.loc, |
2388 | "elemental function call with polymorphic result"); |
2389 | return {}; |
2390 | } |
2391 | |
2392 | bool resultMayRequireFinalization(CallContext &callContext) const { |
2393 | std::optional<Fortran::evaluate::DynamicType> retTy = |
2394 | caller.getCallDescription().proc().GetType(); |
2395 | if (!retTy) |
2396 | return false; |
2397 | |
2398 | if (retTy->IsPolymorphic() || retTy->IsUnlimitedPolymorphic()) |
2399 | fir::emitFatalError( |
2400 | callContext.loc, |
2401 | "elemental function call with [unlimited-]polymorphic result"); |
2402 | |
2403 | if (retTy->category() == Fortran::common::TypeCategory::Derived) { |
2404 | const Fortran::semantics::DerivedTypeSpec &typeSpec = |
2405 | retTy->GetDerivedTypeSpec(); |
2406 | return Fortran::semantics::IsFinalizable(typeSpec); |
2407 | } |
2408 | |
2409 | return false; |
2410 | } |
2411 | |
2412 | private: |
2413 | Fortran::lower::CallerInterface &caller; |
2414 | mlir::FunctionType callSiteType; |
2415 | }; |
2416 | |
2417 | class ElementalIntrinsicCallBuilder |
2418 | : public ElementalCallBuilder<ElementalIntrinsicCallBuilder> { |
2419 | public: |
2420 | ElementalIntrinsicCallBuilder( |
2421 | const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2422 | const fir::IntrinsicHandlerEntry &intrinsicEntry, bool isFunction) |
2423 | : intrinsic{intrinsic}, intrinsicEntry{intrinsicEntry}, |
2424 | isFunction{isFunction} {} |
2425 | std::optional<hlfir::Entity> |
2426 | genElementalKernel(Fortran::lower::PreparedActualArguments &loweredActuals, |
2427 | CallContext &callContext) { |
2428 | return genHLFIRIntrinsicRefCore(loweredActuals, intrinsic, intrinsicEntry, |
2429 | callContext); |
2430 | } |
2431 | // Elemental intrinsic functions cannot modify their arguments. |
2432 | bool argMayBeModifiedByCall(int) const { return !isFunction; } |
2433 | bool canLoadActualArgumentBeforeLoop(int) const { |
2434 | // Elemental intrinsic functions never need the actual addresses |
2435 | // of their arguments. |
2436 | return isFunction; |
2437 | } |
2438 | |
2439 | mlir::Value computeDynamicCharacterResultLength( |
2440 | Fortran::lower::PreparedActualArguments &loweredActuals, |
2441 | CallContext &callContext) { |
2442 | if (intrinsic) |
2443 | if (intrinsic->name == "adjustr"|| intrinsic->name == "adjustl"|| |
2444 | intrinsic->name == "merge") |
2445 | return loweredActuals[0].value().genCharLength( |
2446 | callContext.loc, callContext.getBuilder()); |
2447 | // Character MIN/MAX is the min/max of the arguments length that are |
2448 | // present. |
2449 | TODO(callContext.loc, |
2450 | "compute elemental character min/max function result length in HLFIR"); |
2451 | } |
2452 | |
2453 | mlir::Value getPolymorphicResultMold( |
2454 | Fortran::lower::PreparedActualArguments &loweredActuals, |
2455 | CallContext &callContext) { |
2456 | if (!intrinsic) |
2457 | return {}; |
2458 | |
2459 | if (intrinsic->name == "merge") { |
2460 | // MERGE seems to be the only elemental function that can produce |
2461 | // polymorphic result. The MERGE's result is polymorphic iff |
2462 | // both TSOURCE and FSOURCE are polymorphic, and they also must have |
2463 | // the same declared and dynamic types. So any of them can be used |
2464 | // for the mold. |
2465 | assert(!loweredActuals.empty()); |
2466 | return loweredActuals.front()->getPolymorphicMold(callContext.loc); |
2467 | } |
2468 | |
2469 | return {}; |
2470 | } |
2471 | |
2472 | bool resultMayRequireFinalization( |
2473 | [[maybe_unused]] CallContext &callContext) const { |
2474 | // FIXME: need access to the CallerInterface's return type |
2475 | // to check if the result may need finalization (e.g. the result |
2476 | // of MERGE). |
2477 | return false; |
2478 | } |
2479 | |
2480 | private: |
2481 | const Fortran::evaluate::SpecificIntrinsic *intrinsic; |
2482 | fir::IntrinsicHandlerEntry intrinsicEntry; |
2483 | const bool isFunction; |
2484 | }; |
2485 | } // namespace |
2486 | |
2487 | static std::optional<mlir::Value> |
2488 | genIsPresentIfArgMaybeAbsent(mlir::Location loc, hlfir::Entity actual, |
2489 | const Fortran::lower::SomeExpr &expr, |
2490 | CallContext &callContext, |
2491 | bool passAsAllocatableOrPointer) { |
2492 | if (!Fortran::evaluate::MayBePassedAsAbsentOptional(expr)) |
2493 | return std::nullopt; |
2494 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2495 | if (!passAsAllocatableOrPointer && |
2496 | Fortran::evaluate::IsAllocatableOrPointerObject(expr)) { |
2497 | // Passing Allocatable/Pointer to non-pointer/non-allocatable OPTIONAL. |
2498 | // Fortran 2018 15.5.2.12 point 1: If unallocated/disassociated, it is |
2499 | // as if the argument was absent. The main care here is to not do a |
2500 | // copy-in/copy-out because the temp address, even though pointing to a |
2501 | // null size storage, would not be a nullptr and therefore the argument |
2502 | // would not be considered absent on the callee side. Note: if the |
2503 | // allocatable/pointer is also optional, it cannot be absent as per |
2504 | // 15.5.2.12 point 7. and 8. We rely on this to un-conditionally read |
2505 | // the allocatable/pointer descriptor here. |
2506 | mlir::Value addr = genVariableRawAddress(loc, builder, actual); |
2507 | return builder.genIsNotNullAddr(loc, addr); |
2508 | } |
2509 | // TODO: what if passing allocatable target to optional intent(in) pointer? |
2510 | // May fall into the category above if the allocatable is not optional. |
2511 | |
2512 | // Passing an optional to an optional. |
2513 | return builder.create<fir::IsPresentOp>(loc, builder.getI1Type(), actual) |
2514 | .getResult(); |
2515 | } |
2516 | |
2517 | // Lower a reference to an elemental intrinsic procedure with array arguments |
2518 | // and custom optional handling |
2519 | static std::optional<hlfir::EntityWithAttributes> |
2520 | genCustomElementalIntrinsicRef( |
2521 | const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2522 | CallContext &callContext) { |
2523 | assert(callContext.isElementalProcWithArrayArgs() && |
2524 | "Use genCustomIntrinsicRef for scalar calls"); |
2525 | mlir::Location loc = callContext.loc; |
2526 | auto &converter = callContext.converter; |
2527 | Fortran::lower::PreparedActualArguments operands; |
2528 | assert(intrinsic && Fortran::lower::intrinsicRequiresCustomOptionalHandling( |
2529 | callContext.procRef, *intrinsic, converter)); |
2530 | |
2531 | // callback for optional arguments |
2532 | auto prepareOptionalArg = [&](const Fortran::lower::SomeExpr &expr) { |
2533 | hlfir::EntityWithAttributes actual = Fortran::lower::convertExprToHLFIR( |
2534 | loc, converter, expr, callContext.symMap, callContext.stmtCtx); |
2535 | std::optional<mlir::Value> isPresent = |
2536 | genIsPresentIfArgMaybeAbsent(loc, actual, expr, callContext, |
2537 | /*passAsAllocatableOrPointer=*/false); |
2538 | operands.emplace_back( |
2539 | Fortran::lower::PreparedActualArgument{actual, isPresent}); |
2540 | }; |
2541 | |
2542 | // callback for non-optional arguments |
2543 | auto prepareOtherArg = [&](const Fortran::lower::SomeExpr &expr, |
2544 | fir::LowerIntrinsicArgAs lowerAs) { |
2545 | hlfir::EntityWithAttributes actual = Fortran::lower::convertExprToHLFIR( |
2546 | loc, converter, expr, callContext.symMap, callContext.stmtCtx); |
2547 | operands.emplace_back(Fortran::lower::PreparedActualArgument{ |
2548 | actual, /*isPresent=*/std::nullopt}); |
2549 | }; |
2550 | |
2551 | Fortran::lower::prepareCustomIntrinsicArgument( |
2552 | callContext.procRef, *intrinsic, callContext.resultType, |
2553 | prepareOptionalArg, prepareOtherArg, converter); |
2554 | |
2555 | std::optional<fir::IntrinsicHandlerEntry> intrinsicEntry = |
2556 | fir::lookupIntrinsicHandler(callContext.getBuilder(), |
2557 | callContext.getProcedureName(), |
2558 | callContext.resultType); |
2559 | assert(intrinsicEntry.has_value() && |
2560 | "intrinsic with custom handling for OPTIONAL arguments must have " |
2561 | "lowering entries"); |
2562 | // All of the custom intrinsic elementals with custom handling are pure |
2563 | // functions |
2564 | return ElementalIntrinsicCallBuilder{intrinsic, *intrinsicEntry, |
2565 | /*isFunction=*/true} |
2566 | .genElementalCall(operands, /*isImpure=*/false, callContext); |
2567 | } |
2568 | |
2569 | // Lower a reference to an intrinsic procedure with custom optional handling |
2570 | static std::optional<hlfir::EntityWithAttributes> |
2571 | genCustomIntrinsicRef(const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2572 | CallContext &callContext) { |
2573 | assert(!callContext.isElementalProcWithArrayArgs() && |
2574 | "Needs to be run through ElementalIntrinsicCallBuilder first"); |
2575 | mlir::Location loc = callContext.loc; |
2576 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2577 | auto &converter = callContext.converter; |
2578 | auto &stmtCtx = callContext.stmtCtx; |
2579 | assert(intrinsic && Fortran::lower::intrinsicRequiresCustomOptionalHandling( |
2580 | callContext.procRef, *intrinsic, converter)); |
2581 | Fortran::lower::PreparedActualArguments loweredActuals; |
2582 | |
2583 | // callback for optional arguments |
2584 | auto prepareOptionalArg = [&](const Fortran::lower::SomeExpr &expr) { |
2585 | hlfir::EntityWithAttributes actual = Fortran::lower::convertExprToHLFIR( |
2586 | loc, converter, expr, callContext.symMap, callContext.stmtCtx); |
2587 | mlir::Value isPresent = |
2588 | genIsPresentIfArgMaybeAbsent(loc, actual, expr, callContext, |
2589 | /*passAsAllocatableOrPointer*/ false) |
2590 | .value(); |
2591 | loweredActuals.emplace_back( |
2592 | Fortran::lower::PreparedActualArgument{actual, {isPresent}}); |
2593 | }; |
2594 | |
2595 | // callback for non-optional arguments |
2596 | auto prepareOtherArg = [&](const Fortran::lower::SomeExpr &expr, |
2597 | fir::LowerIntrinsicArgAs lowerAs) { |
2598 | auto getActualFortranElementType = [&]() -> mlir::Type { |
2599 | return hlfir::getFortranElementType(converter.genType(expr)); |
2600 | }; |
2601 | hlfir::EntityWithAttributes actual = Fortran::lower::convertExprToHLFIR( |
2602 | loc, converter, expr, callContext.symMap, callContext.stmtCtx); |
2603 | std::optional<fir::ExtendedValue> exv; |
2604 | switch (lowerAs) { |
2605 | case fir::LowerIntrinsicArgAs::Value: |
2606 | exv = Fortran::lower::convertToValue(loc, converter, actual, stmtCtx); |
2607 | break; |
2608 | case fir::LowerIntrinsicArgAs::Addr: |
2609 | exv = Fortran::lower::convertToAddress(loc, converter, actual, stmtCtx, |
2610 | getActualFortranElementType()); |
2611 | break; |
2612 | case fir::LowerIntrinsicArgAs::Box: |
2613 | exv = Fortran::lower::convertToBox(loc, converter, actual, stmtCtx, |
2614 | getActualFortranElementType()); |
2615 | break; |
2616 | case fir::LowerIntrinsicArgAs::Inquired: |
2617 | exv = Fortran::lower::translateToExtendedValue(loc, builder, actual, |
2618 | stmtCtx); |
2619 | break; |
2620 | } |
2621 | if (!exv) |
2622 | llvm_unreachable("bad switch"); |
2623 | actual = extendedValueToHlfirEntity(loc, builder, exv.value(), |
2624 | "tmp.custom_intrinsic_arg"); |
2625 | loweredActuals.emplace_back(Fortran::lower::PreparedActualArgument{ |
2626 | actual, /*isPresent=*/std::nullopt}); |
2627 | }; |
2628 | |
2629 | Fortran::lower::prepareCustomIntrinsicArgument( |
2630 | callContext.procRef, *intrinsic, callContext.resultType, |
2631 | prepareOptionalArg, prepareOtherArg, converter); |
2632 | |
2633 | return genCustomIntrinsicRefCore(loweredActuals, intrinsic, callContext); |
2634 | } |
2635 | |
2636 | /// Lower an intrinsic procedure reference. |
2637 | /// \p intrinsic is null if this is an intrinsic module procedure that must be |
2638 | /// lowered as if it were an intrinsic module procedure (like C_LOC which is a |
2639 | /// procedure from intrinsic module iso_c_binding). Otherwise, \p intrinsic |
2640 | /// must not be null. |
2641 | |
2642 | static std::optional<hlfir::EntityWithAttributes> |
2643 | genIntrinsicRef(const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2644 | const fir::IntrinsicHandlerEntry &intrinsicEntry, |
2645 | CallContext &callContext) { |
2646 | mlir::Location loc = callContext.loc; |
2647 | Fortran::lower::PreparedActualArguments loweredActuals; |
2648 | const fir::IntrinsicArgumentLoweringRules *argLowering = |
2649 | intrinsicEntry.getArgumentLoweringRules(); |
2650 | for (const auto &arg : llvm::enumerate(callContext.procRef.arguments())) { |
2651 | |
2652 | if (!arg.value()) { |
2653 | // Absent optional. |
2654 | loweredActuals.push_back(std::nullopt); |
2655 | continue; |
2656 | } |
2657 | auto *expr = |
2658 | Fortran::evaluate::UnwrapExpr<Fortran::lower::SomeExpr>(arg.value()); |
2659 | if (!expr) { |
2660 | // TYPE(*) dummy. They are only allowed as argument of a few intrinsics |
2661 | // that do not take optional arguments: see Fortran 2018 standard C710. |
2662 | const Fortran::evaluate::Symbol *assumedTypeSym = |
2663 | arg.value()->GetAssumedTypeDummy(); |
2664 | if (!assumedTypeSym) |
2665 | fir::emitFatalError(loc, |
2666 | "expected assumed-type symbol as actual argument"); |
2667 | std::optional<fir::FortranVariableOpInterface> var = |
2668 | callContext.symMap.lookupVariableDefinition(*assumedTypeSym); |
2669 | if (!var) |
2670 | fir::emitFatalError(loc, "assumed-type symbol was not lowered"); |
2671 | assert( |
2672 | (!argLowering || |
2673 | !fir::lowerIntrinsicArgumentAs(*argLowering, arg.index()) |
2674 | .handleDynamicOptional) && |
2675 | "TYPE(*) are not expected to appear as optional intrinsic arguments"); |
2676 | loweredActuals.push_back(Fortran::lower::PreparedActualArgument{ |
2677 | hlfir::Entity{*var}, /*isPresent=*/std::nullopt}); |
2678 | continue; |
2679 | } |
2680 | // arguments of bitwise comparison functions may not have nsw flag |
2681 | // even if -fno-wrapv is enabled |
2682 | mlir::arith::IntegerOverflowFlags iofBackup{}; |
2683 | auto isBitwiseComparison = [](const std::string intrinsicName) -> bool { |
2684 | if (intrinsicName == "bge"|| intrinsicName == "bgt"|| |
2685 | intrinsicName == "ble"|| intrinsicName == "blt") |
2686 | return true; |
2687 | return false; |
2688 | }; |
2689 | if (isBitwiseComparison(callContext.getProcedureName())) { |
2690 | iofBackup = callContext.getBuilder().getIntegerOverflowFlags(); |
2691 | callContext.getBuilder().setIntegerOverflowFlags( |
2692 | mlir::arith::IntegerOverflowFlags::none); |
2693 | } |
2694 | auto loweredActual = Fortran::lower::convertExprToHLFIR( |
2695 | loc, callContext.converter, *expr, callContext.symMap, |
2696 | callContext.stmtCtx); |
2697 | if (isBitwiseComparison(callContext.getProcedureName())) |
2698 | callContext.getBuilder().setIntegerOverflowFlags(iofBackup); |
2699 | |
2700 | std::optional<mlir::Value> isPresent; |
2701 | if (argLowering) { |
2702 | fir::ArgLoweringRule argRules = |
2703 | fir::lowerIntrinsicArgumentAs(*argLowering, arg.index()); |
2704 | if (argRules.handleDynamicOptional) |
2705 | isPresent = |
2706 | genIsPresentIfArgMaybeAbsent(loc, loweredActual, *expr, callContext, |
2707 | /*passAsAllocatableOrPointer=*/false); |
2708 | } |
2709 | loweredActuals.push_back( |
2710 | Fortran::lower::PreparedActualArgument{loweredActual, isPresent}); |
2711 | } |
2712 | |
2713 | if (callContext.isElementalProcWithArrayArgs()) { |
2714 | // All intrinsic elemental functions are pure. |
2715 | const bool isFunction = callContext.resultType.has_value(); |
2716 | return ElementalIntrinsicCallBuilder{intrinsic, intrinsicEntry, isFunction} |
2717 | .genElementalCall(loweredActuals, /*isImpure=*/!isFunction, |
2718 | callContext); |
2719 | } |
2720 | std::optional<hlfir::EntityWithAttributes> result = genHLFIRIntrinsicRefCore( |
2721 | loweredActuals, intrinsic, intrinsicEntry, callContext); |
2722 | if (result && mlir::isa<hlfir::ExprType>(result->getType())) { |
2723 | fir::FirOpBuilder *bldr = &callContext.getBuilder(); |
2724 | callContext.stmtCtx.attachCleanup( |
2725 | [=]() { bldr->create<hlfir::DestroyOp>(loc, *result); }); |
2726 | } |
2727 | return result; |
2728 | } |
2729 | |
2730 | static std::optional<hlfir::EntityWithAttributes> |
2731 | genIntrinsicRef(const Fortran::evaluate::SpecificIntrinsic *intrinsic, |
2732 | CallContext &callContext) { |
2733 | mlir::Location loc = callContext.loc; |
2734 | auto &converter = callContext.converter; |
2735 | if (intrinsic && Fortran::lower::intrinsicRequiresCustomOptionalHandling( |
2736 | callContext.procRef, *intrinsic, converter)) { |
2737 | if (callContext.isElementalProcWithArrayArgs()) |
2738 | return genCustomElementalIntrinsicRef(intrinsic, callContext); |
2739 | return genCustomIntrinsicRef(intrinsic, callContext); |
2740 | } |
2741 | std::optional<fir::IntrinsicHandlerEntry> intrinsicEntry = |
2742 | fir::lookupIntrinsicHandler(callContext.getBuilder(), |
2743 | callContext.getProcedureName(), |
2744 | callContext.resultType); |
2745 | if (!intrinsicEntry) |
2746 | fir::crashOnMissingIntrinsic(loc, callContext.getProcedureName()); |
2747 | return genIntrinsicRef(intrinsic, *intrinsicEntry, callContext); |
2748 | } |
2749 | |
2750 | /// Main entry point to lower procedure references, regardless of what they are. |
2751 | static std::optional<hlfir::EntityWithAttributes> |
2752 | genProcedureRef(CallContext &callContext) { |
2753 | mlir::Location loc = callContext.loc; |
2754 | fir::FirOpBuilder &builder = callContext.getBuilder(); |
2755 | if (auto *intrinsic = callContext.procRef.proc().GetSpecificIntrinsic()) |
2756 | return genIntrinsicRef(intrinsic, callContext); |
2757 | // Intercept non BIND(C) module procedure reference that have lowering |
2758 | // handlers defined for there name. Otherwise, lower them as user |
2759 | // procedure calls and expect the implementation to be part of |
2760 | // runtime libraries with the proper name mangling. |
2761 | if (Fortran::lower::isIntrinsicModuleProcRef(callContext.procRef) && |
2762 | !callContext.isBindcCall()) |
2763 | if (std::optional<fir::IntrinsicHandlerEntry> intrinsicEntry = |
2764 | fir::lookupIntrinsicHandler(builder, callContext.getProcedureName(), |
2765 | callContext.resultType)) |
2766 | return genIntrinsicRef(nullptr, *intrinsicEntry, callContext); |
2767 | |
2768 | if (callContext.isStatementFunctionCall()) |
2769 | return genStmtFunctionRef(loc, callContext.converter, callContext.symMap, |
2770 | callContext.stmtCtx, callContext.procRef); |
2771 | |
2772 | Fortran::lower::CallerInterface caller(callContext.procRef, |
2773 | callContext.converter); |
2774 | mlir::FunctionType callSiteType = caller.genFunctionType(); |
2775 | const bool isElemental = callContext.isElementalProcWithArrayArgs(); |
2776 | Fortran::lower::PreparedActualArguments loweredActuals; |
2777 | // Lower the actual arguments |
2778 | for (const Fortran::lower::CallInterface< |
2779 | Fortran::lower::CallerInterface>::PassedEntity &arg : |
2780 | caller.getPassedArguments()) |
2781 | if (const auto *actual = arg.entity) { |
2782 | const auto *expr = actual->UnwrapExpr(); |
2783 | if (!expr) { |
2784 | // TYPE(*) actual argument. |
2785 | const Fortran::evaluate::Symbol *assumedTypeSym = |
2786 | actual->GetAssumedTypeDummy(); |
2787 | if (!assumedTypeSym) |
2788 | fir::emitFatalError( |
2789 | loc, "expected assumed-type symbol as actual argument"); |
2790 | std::optional<fir::FortranVariableOpInterface> var = |
2791 | callContext.symMap.lookupVariableDefinition(*assumedTypeSym); |
2792 | if (!var) |
2793 | fir::emitFatalError(loc, "assumed-type symbol was not lowered"); |
2794 | hlfir::Entity actual{*var}; |
2795 | std::optional<mlir::Value> isPresent; |
2796 | if (arg.isOptional()) { |
2797 | // Passing an optional TYPE(*) to an optional TYPE(*). Note that |
2798 | // TYPE(*) cannot be ALLOCATABLE/POINTER (C709) so there is no |
2799 | // need to cover the case of passing an ALLOCATABLE/POINTER to an |
2800 | // OPTIONAL. |
2801 | isPresent = |
2802 | builder.create<fir::IsPresentOp>(loc, builder.getI1Type(), actual) |
2803 | .getResult(); |
2804 | } |
2805 | loweredActuals.push_back(Fortran::lower::PreparedActualArgument{ |
2806 | hlfir::Entity{*var}, isPresent}); |
2807 | continue; |
2808 | } |
2809 | |
2810 | if (Fortran::evaluate::UnwrapExpr<Fortran::evaluate::NullPointer>( |
2811 | *expr)) { |
2812 | if ((arg.passBy != |
2813 | Fortran::lower::CallerInterface::PassEntityBy::MutableBox) && |
2814 | (arg.passBy != |
2815 | Fortran::lower::CallerInterface::PassEntityBy::BoxProcRef)) { |
2816 | assert( |
2817 | arg.isOptional() && |
2818 | "NULL must be passed only to pointer, allocatable, or OPTIONAL"); |
2819 | // Trying to lower NULL() outside of any context would lead to |
2820 | // trouble. NULL() here is equivalent to not providing the |
2821 | // actual argument. |
2822 | loweredActuals.emplace_back(std::nullopt); |
2823 | continue; |
2824 | } |
2825 | } |
2826 | |
2827 | if (isElemental && !arg.hasValueAttribute() && |
2828 | Fortran::evaluate::IsVariable(*expr) && |
2829 | Fortran::evaluate::HasVectorSubscript(*expr)) { |
2830 | // Vector subscripted arguments are copied in calls, except in elemental |
2831 | // calls without VALUE attribute where Fortran 2018 15.5.2.4 point 21 |
2832 | // does not apply and the address of each element must be passed. |
2833 | hlfir::ElementalAddrOp elementalAddr = |
2834 | Fortran::lower::convertVectorSubscriptedExprToElementalAddr( |
2835 | loc, callContext.converter, *expr, callContext.symMap, |
2836 | callContext.stmtCtx); |
2837 | loweredActuals.emplace_back( |
2838 | Fortran::lower::PreparedActualArgument{elementalAddr}); |
2839 | continue; |
2840 | } |
2841 | |
2842 | auto loweredActual = Fortran::lower::convertExprToHLFIR( |
2843 | loc, callContext.converter, *expr, callContext.symMap, |
2844 | callContext.stmtCtx); |
2845 | std::optional<mlir::Value> isPresent; |
2846 | if (arg.isOptional()) |
2847 | isPresent = genIsPresentIfArgMaybeAbsent( |
2848 | loc, loweredActual, *expr, callContext, |
2849 | arg.passBy == |
2850 | Fortran::lower::CallerInterface::PassEntityBy::MutableBox); |
2851 | |
2852 | loweredActuals.emplace_back( |
2853 | Fortran::lower::PreparedActualArgument{loweredActual, isPresent}); |
2854 | } else { |
2855 | // Optional dummy argument for which there is no actual argument. |
2856 | loweredActuals.emplace_back(std::nullopt); |
2857 | } |
2858 | if (isElemental) { |
2859 | bool isImpure = false; |
2860 | if (const Fortran::semantics::Symbol *procSym = |
2861 | callContext.procRef.proc().GetSymbol()) |
2862 | isImpure = !Fortran::semantics::IsPureProcedure(*procSym); |
2863 | return ElementalUserCallBuilder{caller, callSiteType}.genElementalCall( |
2864 | loweredActuals, isImpure, callContext); |
2865 | } |
2866 | return genUserCall(loweredActuals, caller, callSiteType, callContext); |
2867 | } |
2868 | |
2869 | hlfir::Entity Fortran::lower::PreparedActualArgument::getActual( |
2870 | mlir::Location loc, fir::FirOpBuilder &builder) const { |
2871 | if (auto *actualEntity = std::get_if<hlfir::Entity>(&actual)) { |
2872 | if (oneBasedElementalIndices) |
2873 | return hlfir::getElementAt(loc, builder, *actualEntity, |
2874 | *oneBasedElementalIndices); |
2875 | return *actualEntity; |
2876 | } |
2877 | assert(oneBasedElementalIndices && "expect elemental context"); |
2878 | hlfir::ElementalAddrOp elementalAddr = |
2879 | std::get<hlfir::ElementalAddrOp>(actual); |
2880 | mlir::IRMapping mapper; |
2881 | auto alwaysFalse = [](hlfir::ElementalOp) -> bool { return false; }; |
2882 | mlir::Value addr = hlfir::inlineElementalOp( |
2883 | loc, builder, elementalAddr, *oneBasedElementalIndices, mapper, |
2884 | /*mustRecursivelyInline=*/alwaysFalse); |
2885 | assert(elementalAddr.getCleanup().empty() && "no clean-up expected"); |
2886 | elementalAddr.erase(); |
2887 | return hlfir::Entity{addr}; |
2888 | } |
2889 | |
2890 | bool Fortran::lower::isIntrinsicModuleProcRef( |
2891 | const Fortran::evaluate::ProcedureRef &procRef) { |
2892 | const Fortran::semantics::Symbol *symbol = procRef.proc().GetSymbol(); |
2893 | if (!symbol) |
2894 | return false; |
2895 | const Fortran::semantics::Symbol *module = |
2896 | symbol->GetUltimate().owner().GetSymbol(); |
2897 | return module && module->attrs().test(Fortran::semantics::Attr::INTRINSIC); |
2898 | } |
2899 | |
2900 | static bool isInWhereMaskedExpression(fir::FirOpBuilder &builder) { |
2901 | // The MASK of the outer WHERE is not masked itself. |
2902 | mlir::Operation *op = builder.getRegion().getParentOp(); |
2903 | return op && op->getParentOfType<hlfir::WhereOp>(); |
2904 | } |
2905 | |
2906 | std::optional<hlfir::EntityWithAttributes> Fortran::lower::convertCallToHLFIR( |
2907 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
2908 | const evaluate::ProcedureRef &procRef, std::optional<mlir::Type> resultType, |
2909 | Fortran::lower::SymMap &symMap, Fortran::lower::StatementContext &stmtCtx) { |
2910 | auto &builder = converter.getFirOpBuilder(); |
2911 | if (resultType && !procRef.IsElemental() && |
2912 | isInWhereMaskedExpression(builder) && |
2913 | !builder.getRegion().getParentOfType<hlfir::ExactlyOnceOp>()) { |
2914 | // Non elemental calls inside a where-assignment-stmt must be executed |
2915 | // exactly once without mask control. Lower them in a special region so that |
2916 | // this can be enforced whenscheduling forall/where expression evaluations. |
2917 | Fortran::lower::StatementContext localStmtCtx; |
2918 | mlir::Type bogusType = builder.getIndexType(); |
2919 | auto exactlyOnce = builder.create<hlfir::ExactlyOnceOp>(loc, bogusType); |
2920 | mlir::Block *block = builder.createBlock(&exactlyOnce.getBody()); |
2921 | builder.setInsertionPointToStart(block); |
2922 | CallContext callContext(procRef, resultType, loc, converter, symMap, |
2923 | localStmtCtx); |
2924 | std::optional<hlfir::EntityWithAttributes> res = |
2925 | genProcedureRef(callContext); |
2926 | assert(res.has_value() && "must be a function"); |
2927 | auto yield = builder.create<hlfir::YieldOp>(loc, *res); |
2928 | Fortran::lower::genCleanUpInRegionIfAny(loc, builder, yield.getCleanup(), |
2929 | localStmtCtx); |
2930 | builder.setInsertionPointAfter(exactlyOnce); |
2931 | exactlyOnce->getResult(0).setType(res->getType()); |
2932 | if (hlfir::isFortranValue(exactlyOnce.getResult())) |
2933 | return hlfir::EntityWithAttributes{exactlyOnce.getResult()}; |
2934 | // Create hlfir.declare for the result to satisfy |
2935 | // hlfir::EntityWithAttributes requirements. |
2936 | auto [exv, cleanup] = hlfir::translateToExtendedValue( |
2937 | loc, builder, hlfir::Entity{exactlyOnce}); |
2938 | assert(!cleanup && "resut is a variable"); |
2939 | return hlfir::genDeclare(loc, builder, exv, ".func.pointer.result", |
2940 | fir::FortranVariableFlagsAttr{}); |
2941 | } |
2942 | CallContext callContext(procRef, resultType, loc, converter, symMap, stmtCtx); |
2943 | return genProcedureRef(callContext); |
2944 | } |
2945 | |
2946 | void Fortran::lower::convertUserDefinedAssignmentToHLFIR( |
2947 | mlir::Location loc, Fortran::lower::AbstractConverter &converter, |
2948 | const evaluate::ProcedureRef &procRef, hlfir::Entity lhs, hlfir::Entity rhs, |
2949 | Fortran::lower::SymMap &symMap) { |
2950 | Fortran::lower::StatementContext definedAssignmentContext; |
2951 | CallContext callContext(procRef, /*resultType=*/std::nullopt, loc, converter, |
2952 | symMap, definedAssignmentContext); |
2953 | Fortran::lower::CallerInterface caller(procRef, converter); |
2954 | mlir::FunctionType callSiteType = caller.genFunctionType(); |
2955 | PreparedActualArgument preparedLhs{lhs, /*isPresent=*/std::nullopt}; |
2956 | PreparedActualArgument preparedRhs{rhs, /*isPresent=*/std::nullopt}; |
2957 | PreparedActualArguments loweredActuals{preparedLhs, preparedRhs}; |
2958 | genUserCall(loweredActuals, caller, callSiteType, callContext); |
2959 | return; |
2960 | } |
2961 |
Definitions
- useHlfirIntrinsicOps
- tempResultName
- toExtendedValue
- genRecordCPtrValueArg
- findHostAssocTuplePos
- mustCastFuncOpToCopeWithImplicitInterfaceMismatch
- readDim3Value
- remapActualToDummyDescriptor
- remapActualToDummyDescriptors
- genStmtFunctionRef
- CallContext
- CallContext
- getBuilder
- getProcedureName
- isElementalProcWithArrayArgs
- isStatementFunctionCall
- isBindcCall
- extendedValueToHlfirEntity
- CallCleanUp
- CopyIn
- genCleanUp
- ExprAssociate
- genCleanUp
- genCleanUp
- PreparedDummyArgument
- pushCopyInCleanUp
- pushExprAssociateCleanUp
- pushExprAssociateCleanUp
- ConditionallyPreparedDummy
- ConditionallyPreparedDummy
- getIfResulTypes
- genThenResult
- genElseResult
- getPreparedDummy
- fixProcedureDummyMismatch
- getZeroLowerBounds
- isSimplyContiguous
- isParameterObjectOrSubObject
- preparePresentUserCallActualArgument
- prepareUserCallActualArgument
- prepareProcedurePointerActualArgument
- prepareUserCallArguments
- genUserCall
- genOptionalValue
- genOptionalAddr
- genOptionalBox
- genCustomIntrinsicRefCore
- genIntrinsicRefCore
- genHLFIRIntrinsicRefCore
- ElementalCallBuilder
- genElementalCall
- impl
- ElementalUserCallBuilder
- ElementalUserCallBuilder
- genElementalKernel
- argMayBeModifiedByCall
- canLoadActualArgumentBeforeLoop
- computeDynamicCharacterResultLength
- getPolymorphicResultMold
- resultMayRequireFinalization
- ElementalIntrinsicCallBuilder
- ElementalIntrinsicCallBuilder
- genElementalKernel
- argMayBeModifiedByCall
- canLoadActualArgumentBeforeLoop
- computeDynamicCharacterResultLength
- getPolymorphicResultMold
- resultMayRequireFinalization
- genIsPresentIfArgMaybeAbsent
- genCustomElementalIntrinsicRef
- genCustomIntrinsicRef
- genIntrinsicRef
- genIntrinsicRef
- genProcedureRef
Learn to use CMake with our Intro Training
Find out more