1 | //===-- Target.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/Optimizer/CodeGen/Target.h" |
14 | #include "flang/Optimizer/Builder/Todo.h" |
15 | #include "flang/Optimizer/Dialect/FIRType.h" |
16 | #include "flang/Optimizer/Dialect/Support/KindMapping.h" |
17 | #include "flang/Optimizer/Support/FatalError.h" |
18 | #include "flang/Optimizer/Support/Utils.h" |
19 | #include "mlir/IR/BuiltinTypes.h" |
20 | #include "mlir/IR/TypeRange.h" |
21 | #include "llvm/ADT/TypeSwitch.h" |
22 | |
23 | #define DEBUG_TYPE "flang-codegen-target" |
24 | |
25 | using namespace fir; |
26 | |
27 | namespace fir::details { |
28 | llvm::StringRef Attributes::getIntExtensionAttrName() const { |
29 | // The attribute names are available via LLVM dialect interfaces |
30 | // like getZExtAttrName(), getByValAttrName(), etc., so we'd better |
31 | // use them than literals. |
32 | if (isZeroExt()) |
33 | return "llvm.zeroext" ; |
34 | else if (isSignExt()) |
35 | return "llvm.signext" ; |
36 | return {}; |
37 | } |
38 | } // namespace fir::details |
39 | |
40 | // Reduce a REAL/float type to the floating point semantics. |
41 | static const llvm::fltSemantics &floatToSemantics(const KindMapping &kindMap, |
42 | mlir::Type type) { |
43 | assert(isa_real(type)); |
44 | return mlir::cast<mlir::FloatType>(type).getFloatSemantics(); |
45 | } |
46 | |
47 | static void typeTodo(const llvm::fltSemantics *sem, mlir::Location loc, |
48 | const std::string &context) { |
49 | if (sem == &llvm::APFloat::IEEEhalf()) { |
50 | TODO(loc, "COMPLEX(KIND=2): for " + context + " type" ); |
51 | } else if (sem == &llvm::APFloat::BFloat()) { |
52 | TODO(loc, "COMPLEX(KIND=3): " + context + " type" ); |
53 | } else if (sem == &llvm::APFloat::x87DoubleExtended()) { |
54 | TODO(loc, "COMPLEX(KIND=10): " + context + " type" ); |
55 | } else { |
56 | TODO(loc, "complex for this precision for " + context + " type" ); |
57 | } |
58 | } |
59 | |
60 | namespace { |
61 | template <typename S> |
62 | struct GenericTarget : public CodeGenSpecifics { |
63 | using CodeGenSpecifics::CodeGenSpecifics; |
64 | using AT = CodeGenSpecifics::Attributes; |
65 | |
66 | mlir::Type complexMemoryType(mlir::Type eleTy) const override { |
67 | assert(fir::isa_real(eleTy)); |
68 | // Use a type that will be translated into LLVM as: |
69 | // { t, t } struct of 2 eleTy |
70 | return mlir::TupleType::get(eleTy.getContext(), |
71 | mlir::TypeRange{eleTy, eleTy}); |
72 | } |
73 | |
74 | mlir::Type boxcharMemoryType(mlir::Type eleTy) const override { |
75 | auto idxTy = mlir::IntegerType::get(eleTy.getContext(), S::defaultWidth); |
76 | auto ptrTy = fir::ReferenceType::get(eleTy); |
77 | // Use a type that will be translated into LLVM as: |
78 | // { t*, index } |
79 | return mlir::TupleType::get(eleTy.getContext(), |
80 | mlir::TypeRange{ptrTy, idxTy}); |
81 | } |
82 | |
83 | Marshalling boxcharArgumentType(mlir::Type eleTy) const override { |
84 | CodeGenSpecifics::Marshalling marshal; |
85 | auto idxTy = mlir::IntegerType::get(eleTy.getContext(), S::defaultWidth); |
86 | auto ptrTy = fir::ReferenceType::get(eleTy); |
87 | marshal.emplace_back(ptrTy, AT{}); |
88 | // Characters are passed in a split format with all pointers first (in the |
89 | // declared position) and all LEN arguments appended after all of the dummy |
90 | // arguments. |
91 | // NB: Other conventions/ABIs can/should be supported via options. |
92 | marshal.emplace_back(idxTy, AT{/*alignment=*/0, /*byval=*/false, |
93 | /*sret=*/false, /*append=*/true}); |
94 | return marshal; |
95 | } |
96 | |
97 | CodeGenSpecifics::Marshalling |
98 | structArgumentType(mlir::Location loc, fir::RecordType, |
99 | const Marshalling &) const override { |
100 | TODO(loc, "passing VALUE BIND(C) derived type for this target" ); |
101 | } |
102 | |
103 | CodeGenSpecifics::Marshalling |
104 | structReturnType(mlir::Location loc, fir::RecordType ty) const override { |
105 | TODO(loc, "returning BIND(C) derived type for this target" ); |
106 | } |
107 | |
108 | CodeGenSpecifics::Marshalling |
109 | integerArgumentType(mlir::Location loc, |
110 | mlir::IntegerType argTy) const override { |
111 | CodeGenSpecifics::Marshalling marshal; |
112 | AT::IntegerExtension intExt = AT::IntegerExtension::None; |
113 | if (argTy.getWidth() < getCIntTypeWidth()) { |
114 | // isSigned() and isUnsigned() branches below are dead code currently. |
115 | // If needed, we can generate calls with signed/unsigned argument types |
116 | // to more precisely match C side (e.g. for Fortran runtime functions |
117 | // with 'unsigned short' arguments). |
118 | if (argTy.isSigned()) |
119 | intExt = AT::IntegerExtension::Sign; |
120 | else if (argTy.isUnsigned()) |
121 | intExt = AT::IntegerExtension::Zero; |
122 | else if (argTy.isSignless()) { |
123 | // Zero extend for 'i1' and sign extend for other types. |
124 | if (argTy.getWidth() == 1) |
125 | intExt = AT::IntegerExtension::Zero; |
126 | else |
127 | intExt = AT::IntegerExtension::Sign; |
128 | } |
129 | } |
130 | |
131 | marshal.emplace_back(argTy, AT{/*alignment=*/0, /*byval=*/false, |
132 | /*sret=*/false, /*append=*/false, |
133 | /*intExt=*/intExt}); |
134 | return marshal; |
135 | } |
136 | |
137 | CodeGenSpecifics::Marshalling |
138 | integerReturnType(mlir::Location loc, |
139 | mlir::IntegerType argTy) const override { |
140 | return integerArgumentType(loc, argTy); |
141 | } |
142 | |
143 | // Width of 'int' type is 32-bits for almost all targets, except |
144 | // for AVR and MSP430 (see TargetInfo initializations |
145 | // in clang/lib/Basic/Targets). |
146 | unsigned char getCIntTypeWidth() const override { return 32; } |
147 | }; |
148 | } // namespace |
149 | |
150 | //===----------------------------------------------------------------------===// |
151 | // i386 (x86 32 bit) linux target specifics. |
152 | //===----------------------------------------------------------------------===// |
153 | |
154 | namespace { |
155 | struct TargetI386 : public GenericTarget<TargetI386> { |
156 | using GenericTarget::GenericTarget; |
157 | |
158 | static constexpr int defaultWidth = 32; |
159 | |
160 | CodeGenSpecifics::Marshalling |
161 | complexArgumentType(mlir::Location, mlir::Type eleTy) const override { |
162 | assert(fir::isa_real(eleTy)); |
163 | CodeGenSpecifics::Marshalling marshal; |
164 | // Use a type that will be translated into LLVM as: |
165 | // { t, t } struct of 2 eleTy, byval, align 4 |
166 | auto structTy = |
167 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}); |
168 | marshal.emplace_back(fir::ReferenceType::get(structTy), |
169 | AT{/*alignment=*/4, /*byval=*/true}); |
170 | return marshal; |
171 | } |
172 | |
173 | CodeGenSpecifics::Marshalling |
174 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
175 | assert(fir::isa_real(eleTy)); |
176 | CodeGenSpecifics::Marshalling marshal; |
177 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
178 | if (sem == &llvm::APFloat::IEEEsingle()) { |
179 | // i64 pack both floats in a 64-bit GPR |
180 | marshal.emplace_back(mlir::IntegerType::get(eleTy.getContext(), 64), |
181 | AT{}); |
182 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
183 | // Use a type that will be translated into LLVM as: |
184 | // { t, t } struct of 2 eleTy, sret, align 4 |
185 | auto structTy = mlir::TupleType::get(eleTy.getContext(), |
186 | mlir::TypeRange{eleTy, eleTy}); |
187 | marshal.emplace_back(fir::ReferenceType::get(structTy), |
188 | AT{/*alignment=*/4, /*byval=*/false, /*sret=*/true}); |
189 | } else { |
190 | typeTodo(sem, loc, "return" ); |
191 | } |
192 | return marshal; |
193 | } |
194 | }; |
195 | } // namespace |
196 | |
197 | //===----------------------------------------------------------------------===// |
198 | // i386 (x86 32 bit) Windows target specifics. |
199 | //===----------------------------------------------------------------------===// |
200 | |
201 | namespace { |
202 | struct TargetI386Win : public GenericTarget<TargetI386Win> { |
203 | using GenericTarget::GenericTarget; |
204 | |
205 | static constexpr int defaultWidth = 32; |
206 | |
207 | CodeGenSpecifics::Marshalling |
208 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
209 | CodeGenSpecifics::Marshalling marshal; |
210 | // Use a type that will be translated into LLVM as: |
211 | // { t, t } struct of 2 eleTy, byval, align 4 |
212 | auto structTy = |
213 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}); |
214 | marshal.emplace_back(fir::ReferenceType::get(structTy), |
215 | AT{/*align=*/4, /*byval=*/true}); |
216 | return marshal; |
217 | } |
218 | |
219 | CodeGenSpecifics::Marshalling |
220 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
221 | CodeGenSpecifics::Marshalling marshal; |
222 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
223 | if (sem == &llvm::APFloat::IEEEsingle()) { |
224 | // i64 pack both floats in a 64-bit GPR |
225 | marshal.emplace_back(mlir::IntegerType::get(eleTy.getContext(), 64), |
226 | AT{}); |
227 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
228 | // Use a type that will be translated into LLVM as: |
229 | // { double, double } struct of 2 double, sret, align 8 |
230 | marshal.emplace_back( |
231 | fir::ReferenceType::get(mlir::TupleType::get( |
232 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
233 | AT{/*align=*/8, /*byval=*/false, /*sret=*/true}); |
234 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
235 | // Use a type that will be translated into LLVM as: |
236 | // { fp128, fp128 } struct of 2 fp128, sret, align 16 |
237 | marshal.emplace_back( |
238 | fir::ReferenceType::get(mlir::TupleType::get( |
239 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
240 | AT{/*align=*/16, /*byval=*/false, /*sret=*/true}); |
241 | } else if (sem == &llvm::APFloat::x87DoubleExtended()) { |
242 | // Use a type that will be translated into LLVM as: |
243 | // { x86_fp80, x86_fp80 } struct of 2 x86_fp80, sret, align 4 |
244 | marshal.emplace_back( |
245 | fir::ReferenceType::get(mlir::TupleType::get( |
246 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
247 | AT{/*align=*/4, /*byval=*/false, /*sret=*/true}); |
248 | } else { |
249 | typeTodo(sem, loc, "return" ); |
250 | } |
251 | return marshal; |
252 | } |
253 | }; |
254 | } // namespace |
255 | |
256 | //===----------------------------------------------------------------------===// |
257 | // x86_64 (x86 64 bit) linux target specifics. |
258 | //===----------------------------------------------------------------------===// |
259 | |
260 | namespace { |
261 | struct TargetX86_64 : public GenericTarget<TargetX86_64> { |
262 | using GenericTarget::GenericTarget; |
263 | |
264 | static constexpr int defaultWidth = 64; |
265 | |
266 | CodeGenSpecifics::Marshalling |
267 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
268 | CodeGenSpecifics::Marshalling marshal; |
269 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
270 | if (sem == &llvm::APFloat::IEEEsingle()) { |
271 | // <2 x t> vector of 2 eleTy |
272 | marshal.emplace_back(fir::VectorType::get(2, eleTy), AT{}); |
273 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
274 | // FIXME: In case of SSE register exhaustion, the ABI here may be |
275 | // incorrect since LLVM may pass the real via register and the imaginary |
276 | // part via the stack while the ABI it should be all in register or all |
277 | // in memory. Register occupancy must be analyzed here. |
278 | // two distinct double arguments |
279 | marshal.emplace_back(eleTy, AT{}); |
280 | marshal.emplace_back(eleTy, AT{}); |
281 | } else if (sem == &llvm::APFloat::x87DoubleExtended()) { |
282 | // Use a type that will be translated into LLVM as: |
283 | // { x86_fp80, x86_fp80 } struct of 2 fp128, byval, align 16 |
284 | marshal.emplace_back( |
285 | fir::ReferenceType::get(mlir::TupleType::get( |
286 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
287 | AT{/*align=*/16, /*byval=*/true}); |
288 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
289 | // Use a type that will be translated into LLVM as: |
290 | // { fp128, fp128 } struct of 2 fp128, byval, align 16 |
291 | marshal.emplace_back( |
292 | fir::ReferenceType::get(mlir::TupleType::get( |
293 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
294 | AT{/*align=*/16, /*byval=*/true}); |
295 | } else { |
296 | typeTodo(sem, loc, "argument" ); |
297 | } |
298 | return marshal; |
299 | } |
300 | |
301 | CodeGenSpecifics::Marshalling |
302 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
303 | CodeGenSpecifics::Marshalling marshal; |
304 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
305 | if (sem == &llvm::APFloat::IEEEsingle()) { |
306 | // <2 x t> vector of 2 eleTy |
307 | marshal.emplace_back(fir::VectorType::get(2, eleTy), AT{}); |
308 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
309 | // Use a type that will be translated into LLVM as: |
310 | // { double, double } struct of 2 double |
311 | marshal.emplace_back(mlir::TupleType::get(eleTy.getContext(), |
312 | mlir::TypeRange{eleTy, eleTy}), |
313 | AT{}); |
314 | } else if (sem == &llvm::APFloat::x87DoubleExtended()) { |
315 | // { x86_fp80, x86_fp80 } |
316 | marshal.emplace_back(mlir::TupleType::get(eleTy.getContext(), |
317 | mlir::TypeRange{eleTy, eleTy}), |
318 | AT{}); |
319 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
320 | // Use a type that will be translated into LLVM as: |
321 | // { fp128, fp128 } struct of 2 fp128, sret, align 16 |
322 | marshal.emplace_back( |
323 | fir::ReferenceType::get(mlir::TupleType::get( |
324 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
325 | AT{/*align=*/16, /*byval=*/false, /*sret=*/true}); |
326 | } else { |
327 | typeTodo(sem, loc, "return" ); |
328 | } |
329 | return marshal; |
330 | } |
331 | |
332 | /// X86-64 argument classes from System V ABI version 1.0 section 3.2.3. |
333 | enum ArgClass { |
334 | Integer = 0, |
335 | SSE, |
336 | SSEUp, |
337 | X87, |
338 | X87Up, |
339 | ComplexX87, |
340 | NoClass, |
341 | Memory |
342 | }; |
343 | |
344 | /// Classify an argument type or a field of an aggregate type argument. |
345 | /// See System V ABI version 1.0 section 3.2.3. |
346 | /// The Lo and Hi class are set to the class of the lower eight eightbytes |
347 | /// and upper eight eightbytes on return. |
348 | /// If this is called for an aggregate field, the caller is responsible to |
349 | /// do the post-merge. |
350 | void classify(mlir::Location loc, mlir::Type type, std::uint64_t byteOffset, |
351 | ArgClass &Lo, ArgClass &Hi) const { |
352 | Hi = Lo = ArgClass::NoClass; |
353 | ArgClass ¤t = byteOffset < 8 ? Lo : Hi; |
354 | // System V AMD64 ABI 3.2.3. version 1.0 |
355 | llvm::TypeSwitch<mlir::Type>(type) |
356 | .template Case<mlir::IntegerType>([&](mlir::IntegerType intTy) { |
357 | if (intTy.getWidth() == 128) |
358 | Hi = Lo = ArgClass::Integer; |
359 | else |
360 | current = ArgClass::Integer; |
361 | }) |
362 | .template Case<mlir::FloatType>([&](mlir::Type floatTy) { |
363 | const auto *sem = &floatToSemantics(kindMap, floatTy); |
364 | if (sem == &llvm::APFloat::x87DoubleExtended()) { |
365 | Lo = ArgClass::X87; |
366 | Hi = ArgClass::X87Up; |
367 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
368 | Lo = ArgClass::SSE; |
369 | Hi = ArgClass::SSEUp; |
370 | } else { |
371 | current = ArgClass::SSE; |
372 | } |
373 | }) |
374 | .template Case<mlir::ComplexType>([&](mlir::ComplexType cmplx) { |
375 | const auto *sem = &floatToSemantics(kindMap, cmplx.getElementType()); |
376 | if (sem == &llvm::APFloat::x87DoubleExtended()) { |
377 | current = ArgClass::ComplexX87; |
378 | } else { |
379 | fir::SequenceType::Shape shape{2}; |
380 | classifyArray(loc, |
381 | fir::SequenceType::get(shape, cmplx.getElementType()), |
382 | byteOffset, Lo, Hi); |
383 | } |
384 | }) |
385 | .template Case<fir::LogicalType>([&](fir::LogicalType logical) { |
386 | if (kindMap.getLogicalBitsize(logical.getFKind()) == 128) |
387 | Hi = Lo = ArgClass::Integer; |
388 | else |
389 | current = ArgClass::Integer; |
390 | }) |
391 | .template Case<fir::CharacterType>( |
392 | [&](fir::CharacterType character) { current = ArgClass::Integer; }) |
393 | .template Case<fir::SequenceType>([&](fir::SequenceType seqTy) { |
394 | // Array component. |
395 | classifyArray(loc, seqTy, byteOffset, Lo, Hi); |
396 | }) |
397 | .template Case<fir::RecordType>([&](fir::RecordType recTy) { |
398 | // Component that is a derived type. |
399 | classifyStruct(loc, recTy, byteOffset, Lo, Hi); |
400 | }) |
401 | .template Case<fir::VectorType>([&](fir::VectorType vecTy) { |
402 | // Previously marshalled SSE eight byte for a previous struct |
403 | // argument. |
404 | auto *sem = fir::isa_real(vecTy.getEleTy()) |
405 | ? &floatToSemantics(kindMap, vecTy.getEleTy()) |
406 | : nullptr; |
407 | // Not expecting to hit this todo in standard code (it would |
408 | // require some vector type extension). |
409 | if (!(sem == &llvm::APFloat::IEEEsingle() && vecTy.getLen() <= 2) && |
410 | !(sem == &llvm::APFloat::IEEEhalf() && vecTy.getLen() <= 4)) |
411 | TODO(loc, "passing vector argument to C by value" ); |
412 | current = SSE; |
413 | }) |
414 | .Default([&](mlir::Type ty) { |
415 | if (fir::conformsWithPassByRef(ty)) |
416 | current = ArgClass::Integer; // Pointers. |
417 | else |
418 | TODO(loc, "unsupported component type for BIND(C), VALUE derived " |
419 | "type argument" ); |
420 | }); |
421 | } |
422 | |
423 | // Classify fields of a derived type starting at \p offset. Returns the new |
424 | // offset. Post-merge is left to the caller. |
425 | std::uint64_t classifyStruct(mlir::Location loc, fir::RecordType recTy, |
426 | std::uint64_t byteOffset, ArgClass &Lo, |
427 | ArgClass &Hi) const { |
428 | for (auto component : recTy.getTypeList()) { |
429 | if (byteOffset > 16) { |
430 | // See 3.2.3 p. 1 and note 15. Note that when the offset is bigger |
431 | // than 16 bytes here, it is not a single _m256 and or _m512 entity |
432 | // that could fit in AVX registers. |
433 | Lo = Hi = ArgClass::Memory; |
434 | return byteOffset; |
435 | } |
436 | mlir::Type compType = component.second; |
437 | auto [compSize, compAlign] = fir::getTypeSizeAndAlignmentOrCrash( |
438 | loc, compType, getDataLayout(), kindMap); |
439 | byteOffset = llvm::alignTo(byteOffset, compAlign); |
440 | ArgClass LoComp, HiComp; |
441 | classify(loc, compType, byteOffset, LoComp, HiComp); |
442 | Lo = mergeClass(Lo, LoComp); |
443 | Hi = mergeClass(Hi, HiComp); |
444 | byteOffset = byteOffset + llvm::alignTo(compSize, compAlign); |
445 | if (Lo == ArgClass::Memory || Hi == ArgClass::Memory) |
446 | return byteOffset; |
447 | } |
448 | return byteOffset; |
449 | } |
450 | |
451 | // Classify fields of a constant size array type starting at \p offset. |
452 | // Returns the new offset. Post-merge is left to the caller. |
453 | void classifyArray(mlir::Location loc, fir::SequenceType seqTy, |
454 | std::uint64_t byteOffset, ArgClass &Lo, |
455 | ArgClass &Hi) const { |
456 | mlir::Type eleTy = seqTy.getEleTy(); |
457 | const std::uint64_t arraySize = seqTy.getConstantArraySize(); |
458 | auto [eleSize, eleAlign] = fir::getTypeSizeAndAlignmentOrCrash( |
459 | loc, eleTy, getDataLayout(), kindMap); |
460 | std::uint64_t eleStorageSize = llvm::alignTo(eleSize, eleAlign); |
461 | for (std::uint64_t i = 0; i < arraySize; ++i) { |
462 | byteOffset = llvm::alignTo(byteOffset, eleAlign); |
463 | if (byteOffset > 16) { |
464 | // See 3.2.3 p. 1 and note 15. Same as in classifyStruct. |
465 | Lo = Hi = ArgClass::Memory; |
466 | return; |
467 | } |
468 | ArgClass LoComp, HiComp; |
469 | classify(loc, eleTy, byteOffset, LoComp, HiComp); |
470 | Lo = mergeClass(accum: Lo, field: LoComp); |
471 | Hi = mergeClass(accum: Hi, field: HiComp); |
472 | byteOffset = byteOffset + eleStorageSize; |
473 | if (Lo == ArgClass::Memory || Hi == ArgClass::Memory) |
474 | return; |
475 | } |
476 | } |
477 | |
478 | // Goes through the previously marshalled arguments and count the |
479 | // register occupancy to check if there are enough registers left. |
480 | bool hasEnoughRegisters(mlir::Location loc, int neededIntRegisters, |
481 | int neededSSERegisters, |
482 | const Marshalling &previousArguments) const { |
483 | int availIntRegisters = 6; |
484 | int availSSERegisters = 8; |
485 | for (auto typeAndAttr : previousArguments) { |
486 | const auto &attr = std::get<Attributes>(typeAndAttr); |
487 | if (attr.isByVal()) |
488 | continue; // Previous argument passed on the stack. |
489 | ArgClass Lo, Hi; |
490 | Lo = Hi = ArgClass::NoClass; |
491 | classify(loc, std::get<mlir::Type>(typeAndAttr), 0, Lo, Hi); |
492 | // post merge is not needed here since previous aggregate arguments |
493 | // were marshalled into simpler arguments. |
494 | if (Lo == ArgClass::Integer) |
495 | --availIntRegisters; |
496 | else if (Lo == SSE) |
497 | --availSSERegisters; |
498 | if (Hi == ArgClass::Integer) |
499 | --availIntRegisters; |
500 | else if (Hi == ArgClass::SSE) |
501 | --availSSERegisters; |
502 | } |
503 | return availSSERegisters >= neededSSERegisters && |
504 | availIntRegisters >= neededIntRegisters; |
505 | } |
506 | |
507 | /// Argument class merging as described in System V ABI 3.2.3 point 4. |
508 | ArgClass mergeClass(ArgClass accum, ArgClass field) const { |
509 | assert((accum != ArgClass::Memory && accum != ArgClass::ComplexX87) && |
510 | "Invalid accumulated classification during merge." ); |
511 | if (accum == field || field == NoClass) |
512 | return accum; |
513 | if (field == ArgClass::Memory) |
514 | return ArgClass::Memory; |
515 | if (accum == NoClass) |
516 | return field; |
517 | if (accum == Integer || field == Integer) |
518 | return ArgClass::Integer; |
519 | if (field == ArgClass::X87 || field == ArgClass::X87Up || |
520 | field == ArgClass::ComplexX87 || accum == ArgClass::X87 || |
521 | accum == ArgClass::X87Up) |
522 | return Memory; |
523 | return SSE; |
524 | } |
525 | |
526 | /// Argument class post merging as described in System V ABI 3.2.3 point 5. |
527 | void postMerge(std::uint64_t byteSize, ArgClass &Lo, ArgClass &Hi) const { |
528 | if (Hi == ArgClass::Memory) |
529 | Lo = ArgClass::Memory; |
530 | if (Hi == ArgClass::X87Up && Lo != ArgClass::X87) |
531 | Lo = ArgClass::Memory; |
532 | if (byteSize > 16 && (Lo != ArgClass::SSE || Hi != ArgClass::SSEUp)) |
533 | Lo = ArgClass::Memory; |
534 | if (Hi == ArgClass::SSEUp && Lo != ArgClass::SSE) |
535 | Hi = SSE; |
536 | } |
537 | |
538 | /// When \p recTy is a one field record type that can be passed |
539 | /// like the field on its own, returns the field type. Returns |
540 | /// a null type otherwise. |
541 | mlir::Type passAsFieldIfOneFieldStruct(fir::RecordType recTy, |
542 | bool allowComplex = false) const { |
543 | auto typeList = recTy.getTypeList(); |
544 | if (typeList.size() != 1) |
545 | return {}; |
546 | mlir::Type fieldType = typeList[0].second; |
547 | if (mlir::isa<mlir::FloatType, mlir::IntegerType, fir::LogicalType>( |
548 | fieldType)) |
549 | return fieldType; |
550 | if (allowComplex && mlir::isa<mlir::ComplexType>(fieldType)) |
551 | return fieldType; |
552 | if (mlir::isa<fir::CharacterType>(fieldType)) { |
553 | // Only CHARACTER(1) are expected in BIND(C) contexts, which is the only |
554 | // contexts where derived type may be passed in registers. |
555 | assert(mlir::cast<fir::CharacterType>(fieldType).getLen() == 1 && |
556 | "fir.type value arg character components must have length 1" ); |
557 | return fieldType; |
558 | } |
559 | // Complex field that needs to be split, or array. |
560 | return {}; |
561 | } |
562 | |
563 | mlir::Type pickLLVMArgType(mlir::Location loc, mlir::MLIRContext *context, |
564 | ArgClass argClass, |
565 | std::uint64_t partByteSize) const { |
566 | if (argClass == ArgClass::SSE) { |
567 | if (partByteSize > 16) |
568 | TODO(loc, "passing struct as a real > 128 bits in register" ); |
569 | // Clang uses vector type when several fp fields are marshalled |
570 | // into a single SSE register (like <n x smallest fp field> ). |
571 | // It should make no difference from an ABI point of view to just |
572 | // select an fp type of the right size, and it makes things simpler |
573 | // here. |
574 | if (partByteSize > 8) |
575 | return mlir::Float128Type::get(context); |
576 | if (partByteSize > 4) |
577 | return mlir::Float64Type::get(context); |
578 | if (partByteSize > 2) |
579 | return mlir::Float32Type::get(context); |
580 | return mlir::Float16Type::get(context); |
581 | } |
582 | assert(partByteSize <= 8 && |
583 | "expect integer part of aggregate argument to fit into eight bytes" ); |
584 | if (partByteSize > 4) |
585 | return mlir::IntegerType::get(context, 64); |
586 | if (partByteSize > 2) |
587 | return mlir::IntegerType::get(context, 32); |
588 | if (partByteSize > 1) |
589 | return mlir::IntegerType::get(context, 16); |
590 | return mlir::IntegerType::get(context, 8); |
591 | } |
592 | |
593 | /// Marshal a derived type passed by value like a C struct. |
594 | CodeGenSpecifics::Marshalling |
595 | structArgumentType(mlir::Location loc, fir::RecordType recTy, |
596 | const Marshalling &previousArguments) const override { |
597 | std::uint64_t byteOffset = 0; |
598 | ArgClass Lo, Hi; |
599 | Lo = Hi = ArgClass::NoClass; |
600 | byteOffset = classifyStruct(loc, recTy, byteOffset, Lo, Hi); |
601 | postMerge(byteSize: byteOffset, Lo, Hi); |
602 | if (Lo == ArgClass::Memory || Lo == ArgClass::X87 || |
603 | Lo == ArgClass::ComplexX87) |
604 | return passOnTheStack(loc, recTy, /*isResult=*/false); |
605 | int neededIntRegisters = 0; |
606 | int neededSSERegisters = 0; |
607 | if (Lo == ArgClass::SSE) |
608 | ++neededSSERegisters; |
609 | else if (Lo == ArgClass::Integer) |
610 | ++neededIntRegisters; |
611 | if (Hi == ArgClass::SSE) |
612 | ++neededSSERegisters; |
613 | else if (Hi == ArgClass::Integer) |
614 | ++neededIntRegisters; |
615 | // C struct should not be split into LLVM registers if LLVM codegen is not |
616 | // able to later assign actual registers to all of them (struct passing is |
617 | // all in registers or all on the stack). |
618 | if (!hasEnoughRegisters(loc, neededIntRegisters, neededSSERegisters, |
619 | previousArguments)) |
620 | return passOnTheStack(loc, recTy, /*isResult=*/false); |
621 | |
622 | if (auto fieldType = passAsFieldIfOneFieldStruct(recTy)) { |
623 | CodeGenSpecifics::Marshalling marshal; |
624 | marshal.emplace_back(fieldType, AT{}); |
625 | return marshal; |
626 | } |
627 | if (Hi == ArgClass::NoClass || Hi == ArgClass::SSEUp) { |
628 | // Pass a single integer or floating point argument. |
629 | mlir::Type lowType = |
630 | pickLLVMArgType(loc, recTy.getContext(), Lo, byteOffset); |
631 | CodeGenSpecifics::Marshalling marshal; |
632 | marshal.emplace_back(lowType, AT{}); |
633 | return marshal; |
634 | } |
635 | // Split into two integer or floating point arguments. |
636 | // Note that for the first argument, this will always pick i64 or f64 which |
637 | // may be bigger than needed if some struct padding ends the first eight |
638 | // byte (e.g. for `{i32, f64}`). It is valid from an X86-64 ABI and |
639 | // semantic point of view, but it may not match the LLVM IR interface clang |
640 | // would produce for the equivalent C code (the assembly will still be |
641 | // compatible). This allows keeping the logic simpler here since it |
642 | // avoids computing the "data" size of the Lo part. |
643 | mlir::Type lowType = pickLLVMArgType(loc, recTy.getContext(), Lo, 8u); |
644 | mlir::Type hiType = |
645 | pickLLVMArgType(loc, recTy.getContext(), Hi, byteOffset - 8u); |
646 | CodeGenSpecifics::Marshalling marshal; |
647 | marshal.emplace_back(lowType, AT{}); |
648 | marshal.emplace_back(hiType, AT{}); |
649 | return marshal; |
650 | } |
651 | |
652 | CodeGenSpecifics::Marshalling |
653 | structReturnType(mlir::Location loc, fir::RecordType recTy) const override { |
654 | std::uint64_t byteOffset = 0; |
655 | ArgClass Lo, Hi; |
656 | Lo = Hi = ArgClass::NoClass; |
657 | byteOffset = classifyStruct(loc, recTy, byteOffset, Lo, Hi); |
658 | mlir::MLIRContext *context = recTy.getContext(); |
659 | postMerge(byteSize: byteOffset, Lo, Hi); |
660 | if (Lo == ArgClass::Memory) |
661 | return passOnTheStack(loc, recTy, /*isResult=*/true); |
662 | |
663 | // Note that X87/ComplexX87 are passed in memory, but returned via %st0 |
664 | // %st1 registers. Here, they are returned as fp80 or {fp80, fp80} by |
665 | // passAsFieldIfOneFieldStruct, and LLVM will use the expected registers. |
666 | |
667 | // Note that {_Complex long double} is not 100% clear from an ABI |
668 | // perspective because the aggregate post merger rules say it should be |
669 | // passed in memory because it is bigger than 2 eight bytes. This has the |
670 | // funny effect of |
671 | // {_Complex long double} return to be dealt with differently than |
672 | // _Complex long double. |
673 | |
674 | if (auto fieldType = |
675 | passAsFieldIfOneFieldStruct(recTy, /*allowComplex=*/true)) { |
676 | if (auto complexType = mlir::dyn_cast<mlir::ComplexType>(fieldType)) |
677 | return complexReturnType(loc, complexType.getElementType()); |
678 | CodeGenSpecifics::Marshalling marshal; |
679 | marshal.emplace_back(fieldType, AT{}); |
680 | return marshal; |
681 | } |
682 | |
683 | if (Hi == ArgClass::NoClass || Hi == ArgClass::SSEUp) { |
684 | // Return a single integer or floating point argument. |
685 | mlir::Type lowType = pickLLVMArgType(loc, context, Lo, byteOffset); |
686 | CodeGenSpecifics::Marshalling marshal; |
687 | marshal.emplace_back(lowType, AT{}); |
688 | return marshal; |
689 | } |
690 | // Will be returned in two different registers. Generate {lowTy, HiTy} for |
691 | // the LLVM IR result type. |
692 | CodeGenSpecifics::Marshalling marshal; |
693 | mlir::Type lowType = pickLLVMArgType(loc, context, Lo, 8u); |
694 | mlir::Type hiType = pickLLVMArgType(loc, context, Hi, byteOffset - 8u); |
695 | marshal.emplace_back(mlir::TupleType::get(context, {lowType, hiType}), |
696 | AT{}); |
697 | return marshal; |
698 | } |
699 | |
700 | /// Marshal an argument that must be passed on the stack. |
701 | CodeGenSpecifics::Marshalling |
702 | passOnTheStack(mlir::Location loc, mlir::Type ty, bool isResult) const { |
703 | CodeGenSpecifics::Marshalling marshal; |
704 | auto sizeAndAlign = |
705 | fir::getTypeSizeAndAlignmentOrCrash(loc, ty, getDataLayout(), kindMap); |
706 | // The stack is always 8 byte aligned (note 14 in 3.2.3). |
707 | unsigned short align = |
708 | std::max(sizeAndAlign.second, static_cast<unsigned short>(8)); |
709 | marshal.emplace_back(fir::ReferenceType::get(ty), |
710 | AT{align, /*byval=*/!isResult, /*sret=*/isResult}); |
711 | return marshal; |
712 | } |
713 | }; |
714 | } // namespace |
715 | |
716 | //===----------------------------------------------------------------------===// |
717 | // x86_64 (x86 64 bit) Windows target specifics. |
718 | //===----------------------------------------------------------------------===// |
719 | |
720 | namespace { |
721 | struct TargetX86_64Win : public GenericTarget<TargetX86_64Win> { |
722 | using GenericTarget::GenericTarget; |
723 | |
724 | static constexpr int defaultWidth = 64; |
725 | |
726 | CodeGenSpecifics::Marshalling |
727 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
728 | CodeGenSpecifics::Marshalling marshal; |
729 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
730 | if (sem == &llvm::APFloat::IEEEsingle()) { |
731 | // i64 pack both floats in a 64-bit GPR |
732 | marshal.emplace_back(mlir::IntegerType::get(eleTy.getContext(), 64), |
733 | AT{}); |
734 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
735 | // Use a type that will be translated into LLVM as: |
736 | // { double, double } struct of 2 double, byval, align 8 |
737 | marshal.emplace_back( |
738 | fir::ReferenceType::get(mlir::TupleType::get( |
739 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
740 | AT{/*align=*/8, /*byval=*/true}); |
741 | } else if (sem == &llvm::APFloat::IEEEquad() || |
742 | sem == &llvm::APFloat::x87DoubleExtended()) { |
743 | // Use a type that will be translated into LLVM as: |
744 | // { t, t } struct of 2 eleTy, byval, align 16 |
745 | marshal.emplace_back( |
746 | fir::ReferenceType::get(mlir::TupleType::get( |
747 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
748 | AT{/*align=*/16, /*byval=*/true}); |
749 | } else { |
750 | typeTodo(sem, loc, "argument" ); |
751 | } |
752 | return marshal; |
753 | } |
754 | |
755 | CodeGenSpecifics::Marshalling |
756 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
757 | CodeGenSpecifics::Marshalling marshal; |
758 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
759 | if (sem == &llvm::APFloat::IEEEsingle()) { |
760 | // i64 pack both floats in a 64-bit GPR |
761 | marshal.emplace_back(mlir::IntegerType::get(eleTy.getContext(), 64), |
762 | AT{}); |
763 | } else if (sem == &llvm::APFloat::IEEEdouble()) { |
764 | // Use a type that will be translated into LLVM as: |
765 | // { double, double } struct of 2 double, sret, align 8 |
766 | marshal.emplace_back( |
767 | fir::ReferenceType::get(mlir::TupleType::get( |
768 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
769 | AT{/*align=*/8, /*byval=*/false, /*sret=*/true}); |
770 | } else if (sem == &llvm::APFloat::IEEEquad() || |
771 | sem == &llvm::APFloat::x87DoubleExtended()) { |
772 | // Use a type that will be translated into LLVM as: |
773 | // { t, t } struct of 2 eleTy, sret, align 16 |
774 | marshal.emplace_back( |
775 | fir::ReferenceType::get(mlir::TupleType::get( |
776 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
777 | AT{/*align=*/16, /*byval=*/false, /*sret=*/true}); |
778 | } else { |
779 | typeTodo(sem, loc, "return" ); |
780 | } |
781 | return marshal; |
782 | } |
783 | }; |
784 | } // namespace |
785 | |
786 | //===----------------------------------------------------------------------===// |
787 | // AArch64 target specifics. |
788 | //===----------------------------------------------------------------------===// |
789 | |
790 | namespace { |
791 | // AArch64 procedure call standard: |
792 | // https://github.com/ARM-software/abi-aa/blob/main/aapcs64/aapcs64.rst#parameter-passing |
793 | struct TargetAArch64 : public GenericTarget<TargetAArch64> { |
794 | using GenericTarget::GenericTarget; |
795 | |
796 | static constexpr int defaultWidth = 64; |
797 | |
798 | CodeGenSpecifics::Marshalling |
799 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
800 | CodeGenSpecifics::Marshalling marshal; |
801 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
802 | if (sem == &llvm::APFloat::IEEEsingle() || |
803 | sem == &llvm::APFloat::IEEEdouble() || |
804 | sem == &llvm::APFloat::IEEEquad()) { |
805 | // [2 x t] array of 2 eleTy |
806 | marshal.emplace_back(fir::SequenceType::get({2}, eleTy), AT{}); |
807 | } else { |
808 | typeTodo(sem, loc, "argument" ); |
809 | } |
810 | return marshal; |
811 | } |
812 | |
813 | CodeGenSpecifics::Marshalling |
814 | integerArgumentType(mlir::Location loc, |
815 | mlir::IntegerType argTy) const override { |
816 | if (argTy.getWidth() < getCIntTypeWidth() && argTy.isSignless()) { |
817 | AT::IntegerExtension intExt; |
818 | if (argTy.getWidth() == 1) { |
819 | // Zero extend for 'i1'. |
820 | intExt = AT::IntegerExtension::Zero; |
821 | } else { |
822 | if (triple.isOSDarwin()) { |
823 | // On Darwin, sign extend. The apple developer guide specifies this as |
824 | // a divergence from the AArch64PCS: |
825 | // https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms#Pass-arguments-to-functions-correctly |
826 | intExt = AT::IntegerExtension::Sign; |
827 | } else { |
828 | // On linux, pass directly and do not extend. |
829 | intExt = AT::IntegerExtension::None; |
830 | } |
831 | } |
832 | CodeGenSpecifics::Marshalling marshal; |
833 | marshal.emplace_back(argTy, AT{/*alignment=*/0, /*byval=*/false, |
834 | /*sret=*/false, /*append=*/false, |
835 | /*intExt=*/intExt}); |
836 | return marshal; |
837 | } |
838 | return GenericTarget::integerArgumentType(loc, argTy); |
839 | } |
840 | |
841 | CodeGenSpecifics::Marshalling |
842 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
843 | CodeGenSpecifics::Marshalling marshal; |
844 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
845 | if (sem == &llvm::APFloat::IEEEsingle() || |
846 | sem == &llvm::APFloat::IEEEdouble() || |
847 | sem == &llvm::APFloat::IEEEquad()) { |
848 | // Use a type that will be translated into LLVM as: |
849 | // { t, t } struct of 2 eleTy |
850 | marshal.emplace_back(mlir::TupleType::get(eleTy.getContext(), |
851 | mlir::TypeRange{eleTy, eleTy}), |
852 | AT{}); |
853 | } else { |
854 | typeTodo(sem, loc, "return" ); |
855 | } |
856 | return marshal; |
857 | } |
858 | |
859 | // Flatten a RecordType::TypeList containing more record types or array type |
860 | static std::optional<std::vector<mlir::Type>> |
861 | flattenTypeList(const RecordType::TypeList &types) { |
862 | std::vector<mlir::Type> flatTypes; |
863 | // The flat list will be at least the same size as the non-flat list. |
864 | flatTypes.reserve(types.size()); |
865 | for (auto [c, type] : types) { |
866 | // Flatten record type |
867 | if (auto recTy = mlir::dyn_cast<RecordType>(type)) { |
868 | auto subTypeList = flattenTypeList(recTy.getTypeList()); |
869 | if (!subTypeList) |
870 | return std::nullopt; |
871 | llvm::copy(*subTypeList, std::back_inserter(flatTypes)); |
872 | continue; |
873 | } |
874 | |
875 | // Flatten array type |
876 | if (auto seqTy = mlir::dyn_cast<SequenceType>(type)) { |
877 | if (seqTy.hasDynamicExtents()) |
878 | return std::nullopt; |
879 | std::size_t n = seqTy.getConstantArraySize(); |
880 | auto eleTy = seqTy.getElementType(); |
881 | // Flatten array of record types |
882 | if (auto recTy = mlir::dyn_cast<RecordType>(eleTy)) { |
883 | auto subTypeList = flattenTypeList(recTy.getTypeList()); |
884 | if (!subTypeList) |
885 | return std::nullopt; |
886 | for (std::size_t i = 0; i < n; ++i) |
887 | llvm::copy(*subTypeList, std::back_inserter(flatTypes)); |
888 | } else { |
889 | std::fill_n(std::back_inserter(flatTypes), |
890 | seqTy.getConstantArraySize(), eleTy); |
891 | } |
892 | continue; |
893 | } |
894 | |
895 | // Other types are already flat |
896 | flatTypes.push_back(type); |
897 | } |
898 | return flatTypes; |
899 | } |
900 | |
901 | // Determine if the type is a Homogenous Floating-point Aggregate (HFA). An |
902 | // HFA is a record type with up to 4 floating-point members of the same type. |
903 | static std::optional<int> usedRegsForHFA(fir::RecordType ty) { |
904 | RecordType::TypeList types = ty.getTypeList(); |
905 | if (types.empty() || types.size() > 4) |
906 | return std::nullopt; |
907 | |
908 | std::optional<std::vector<mlir::Type>> flatTypes = flattenTypeList(types); |
909 | if (!flatTypes || flatTypes->size() > 4) { |
910 | return std::nullopt; |
911 | } |
912 | |
913 | if (!isa_real(flatTypes->front())) { |
914 | return std::nullopt; |
915 | } |
916 | |
917 | return llvm::all_equal(*flatTypes) ? std::optional<int>{flatTypes->size()} |
918 | : std::nullopt; |
919 | } |
920 | |
921 | struct NRegs { |
922 | int n{0}; |
923 | bool isSimd{false}; |
924 | }; |
925 | |
926 | NRegs usedRegsForRecordType(mlir::Location loc, fir::RecordType type) const { |
927 | if (std::optional<int> size = usedRegsForHFA(type)) |
928 | return {.n: *size, .isSimd: true}; |
929 | |
930 | auto [size, align] = fir::getTypeSizeAndAlignmentOrCrash( |
931 | loc, type, getDataLayout(), kindMap); |
932 | |
933 | if (size <= 16) |
934 | return {static_cast<int>((size + 7) / 8), false}; |
935 | |
936 | // Pass on the stack, i.e. no registers used |
937 | return {}; |
938 | } |
939 | |
940 | NRegs usedRegsForType(mlir::Location loc, mlir::Type type) const { |
941 | return llvm::TypeSwitch<mlir::Type, NRegs>(type) |
942 | .Case<mlir::IntegerType>([&](auto intTy) { |
943 | return intTy.getWidth() == 128 ? NRegs{2, false} : NRegs{1, false}; |
944 | }) |
945 | .Case<mlir::FloatType>([&](auto) { return NRegs{1, true}; }) |
946 | .Case<mlir::ComplexType>([&](auto) { return NRegs{2, true}; }) |
947 | .Case<fir::LogicalType>([&](auto) { return NRegs{1, false}; }) |
948 | .Case<fir::CharacterType>([&](auto) { return NRegs{1, false}; }) |
949 | .Case<fir::SequenceType>([&](auto ty) { |
950 | assert(ty.getShape().size() == 1 && |
951 | "invalid array dimensions in BIND(C)" ); |
952 | NRegs nregs = usedRegsForType(loc, ty.getEleTy()); |
953 | nregs.n *= ty.getShape()[0]; |
954 | return nregs; |
955 | }) |
956 | .Case<fir::RecordType>( |
957 | [&](auto ty) { return usedRegsForRecordType(loc, ty); }) |
958 | .Case<fir::VectorType>([&](auto) { |
959 | TODO(loc, "passing vector argument to C by value is not supported" ); |
960 | return NRegs{}; |
961 | }) |
962 | .Default([&](auto ty) { |
963 | if (fir::conformsWithPassByRef(ty)) |
964 | return NRegs{1, false}; // Pointers take 1 integer register |
965 | TODO(loc, "unsupported component type for BIND(C), VALUE derived " |
966 | "type argument" ); |
967 | return NRegs{}; |
968 | }); |
969 | } |
970 | |
971 | bool hasEnoughRegisters(mlir::Location loc, fir::RecordType type, |
972 | const Marshalling &previousArguments) const { |
973 | int availIntRegisters = 8; |
974 | int availSIMDRegisters = 8; |
975 | |
976 | // Check previous arguments to see how many registers are used already |
977 | for (auto [type, attr] : previousArguments) { |
978 | if (availIntRegisters <= 0 || availSIMDRegisters <= 0) |
979 | break; |
980 | |
981 | if (attr.isByVal()) |
982 | continue; // Previous argument passed on the stack |
983 | |
984 | NRegs nregs = usedRegsForType(loc, type); |
985 | if (nregs.isSimd) |
986 | availSIMDRegisters -= nregs.n; |
987 | else |
988 | availIntRegisters -= nregs.n; |
989 | } |
990 | |
991 | NRegs nregs = usedRegsForRecordType(loc, type); |
992 | |
993 | if (nregs.isSimd) |
994 | return nregs.n <= availSIMDRegisters; |
995 | |
996 | return nregs.n <= availIntRegisters; |
997 | } |
998 | |
999 | CodeGenSpecifics::Marshalling |
1000 | passOnTheStack(mlir::Location loc, mlir::Type ty, bool isResult) const { |
1001 | CodeGenSpecifics::Marshalling marshal; |
1002 | auto sizeAndAlign = |
1003 | fir::getTypeSizeAndAlignmentOrCrash(loc, ty, getDataLayout(), kindMap); |
1004 | // The stack is always 8 byte aligned |
1005 | unsigned short align = |
1006 | std::max(sizeAndAlign.second, static_cast<unsigned short>(8)); |
1007 | marshal.emplace_back(fir::ReferenceType::get(ty), |
1008 | AT{align, /*byval=*/!isResult, /*sret=*/isResult}); |
1009 | return marshal; |
1010 | } |
1011 | |
1012 | CodeGenSpecifics::Marshalling |
1013 | structType(mlir::Location loc, fir::RecordType type, bool isResult) const { |
1014 | NRegs nregs = usedRegsForRecordType(loc, type); |
1015 | |
1016 | // If the type needs no registers it must need to be passed on the stack |
1017 | if (nregs.n == 0) |
1018 | return passOnTheStack(loc, type, isResult); |
1019 | |
1020 | CodeGenSpecifics::Marshalling marshal; |
1021 | |
1022 | mlir::Type pcsType; |
1023 | if (nregs.isSimd) { |
1024 | pcsType = type; |
1025 | } else { |
1026 | pcsType = fir::SequenceType::get( |
1027 | nregs.n, mlir::IntegerType::get(type.getContext(), 64)); |
1028 | } |
1029 | |
1030 | marshal.emplace_back(pcsType, AT{}); |
1031 | return marshal; |
1032 | } |
1033 | |
1034 | CodeGenSpecifics::Marshalling |
1035 | structArgumentType(mlir::Location loc, fir::RecordType ty, |
1036 | const Marshalling &previousArguments) const override { |
1037 | if (!hasEnoughRegisters(loc, ty, previousArguments)) { |
1038 | return passOnTheStack(loc, ty, /*isResult=*/false); |
1039 | } |
1040 | |
1041 | return structType(loc, ty, /*isResult=*/false); |
1042 | } |
1043 | |
1044 | CodeGenSpecifics::Marshalling |
1045 | structReturnType(mlir::Location loc, fir::RecordType ty) const override { |
1046 | return structType(loc, ty, /*isResult=*/true); |
1047 | } |
1048 | }; |
1049 | } // namespace |
1050 | |
1051 | //===----------------------------------------------------------------------===// |
1052 | // PPC (AIX 32 bit) target specifics. |
1053 | //===----------------------------------------------------------------------===// |
1054 | namespace { |
1055 | struct TargetPPC : public GenericTarget<TargetPPC> { |
1056 | using GenericTarget::GenericTarget; |
1057 | |
1058 | static constexpr int defaultWidth = 32; |
1059 | |
1060 | CodeGenSpecifics::Marshalling |
1061 | complexArgumentType(mlir::Location, mlir::Type eleTy) const override { |
1062 | CodeGenSpecifics::Marshalling marshal; |
1063 | // two distinct element type arguments (re, im) |
1064 | marshal.emplace_back(eleTy, AT{}); |
1065 | marshal.emplace_back(eleTy, AT{}); |
1066 | return marshal; |
1067 | } |
1068 | |
1069 | CodeGenSpecifics::Marshalling |
1070 | complexReturnType(mlir::Location, mlir::Type eleTy) const override { |
1071 | CodeGenSpecifics::Marshalling marshal; |
1072 | // Use a type that will be translated into LLVM as: |
1073 | // { t, t } struct of 2 element type |
1074 | marshal.emplace_back( |
1075 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}), |
1076 | AT{}); |
1077 | return marshal; |
1078 | } |
1079 | }; |
1080 | } // namespace |
1081 | |
1082 | //===----------------------------------------------------------------------===// |
1083 | // PPC64 (AIX 64 bit) target specifics. |
1084 | //===----------------------------------------------------------------------===// |
1085 | |
1086 | namespace { |
1087 | struct TargetPPC64 : public GenericTarget<TargetPPC64> { |
1088 | using GenericTarget::GenericTarget; |
1089 | |
1090 | static constexpr int defaultWidth = 64; |
1091 | |
1092 | CodeGenSpecifics::Marshalling |
1093 | complexArgumentType(mlir::Location, mlir::Type eleTy) const override { |
1094 | CodeGenSpecifics::Marshalling marshal; |
1095 | // two distinct element type arguments (re, im) |
1096 | marshal.emplace_back(eleTy, AT{}); |
1097 | marshal.emplace_back(eleTy, AT{}); |
1098 | return marshal; |
1099 | } |
1100 | |
1101 | CodeGenSpecifics::Marshalling |
1102 | complexReturnType(mlir::Location, mlir::Type eleTy) const override { |
1103 | CodeGenSpecifics::Marshalling marshal; |
1104 | // Use a type that will be translated into LLVM as: |
1105 | // { t, t } struct of 2 element type |
1106 | marshal.emplace_back( |
1107 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}), |
1108 | AT{}); |
1109 | return marshal; |
1110 | } |
1111 | |
1112 | CodeGenSpecifics::Marshalling |
1113 | structType(mlir::Location loc, fir::RecordType ty, bool isResult) const { |
1114 | CodeGenSpecifics::Marshalling marshal; |
1115 | auto sizeAndAlign{ |
1116 | fir::getTypeSizeAndAlignmentOrCrash(loc, ty, getDataLayout(), kindMap)}; |
1117 | unsigned short align{ |
1118 | std::max(sizeAndAlign.second, static_cast<unsigned short>(8))}; |
1119 | marshal.emplace_back(fir::ReferenceType::get(ty), |
1120 | AT{align, /*byval*/ !isResult, /*sret*/ isResult}); |
1121 | return marshal; |
1122 | } |
1123 | |
1124 | CodeGenSpecifics::Marshalling |
1125 | structArgumentType(mlir::Location loc, fir::RecordType ty, |
1126 | const Marshalling &previousArguments) const override { |
1127 | return structType(loc, ty, false); |
1128 | } |
1129 | |
1130 | CodeGenSpecifics::Marshalling |
1131 | structReturnType(mlir::Location loc, fir::RecordType ty) const override { |
1132 | return structType(loc, ty, true); |
1133 | } |
1134 | }; |
1135 | } // namespace |
1136 | |
1137 | //===----------------------------------------------------------------------===// |
1138 | // PPC64le linux target specifics. |
1139 | //===----------------------------------------------------------------------===// |
1140 | |
1141 | namespace { |
1142 | struct TargetPPC64le : public GenericTarget<TargetPPC64le> { |
1143 | using GenericTarget::GenericTarget; |
1144 | |
1145 | static constexpr int defaultWidth{64}; |
1146 | |
1147 | CodeGenSpecifics::Marshalling |
1148 | complexArgumentType(mlir::Location, mlir::Type eleTy) const override { |
1149 | CodeGenSpecifics::Marshalling marshal; |
1150 | // two distinct element type arguments (re, im) |
1151 | marshal.emplace_back(eleTy, AT{}); |
1152 | marshal.emplace_back(eleTy, AT{}); |
1153 | return marshal; |
1154 | } |
1155 | |
1156 | CodeGenSpecifics::Marshalling |
1157 | complexReturnType(mlir::Location, mlir::Type eleTy) const override { |
1158 | CodeGenSpecifics::Marshalling marshal; |
1159 | // Use a type that will be translated into LLVM as: |
1160 | // { t, t } struct of 2 element type |
1161 | marshal.emplace_back( |
1162 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}), |
1163 | AT{}); |
1164 | return marshal; |
1165 | } |
1166 | |
1167 | unsigned getElemWidth(mlir::Type ty) const { |
1168 | unsigned width{}; |
1169 | llvm::TypeSwitch<mlir::Type>(ty) |
1170 | .template Case<mlir::ComplexType>([&](mlir::ComplexType cmplx) { |
1171 | auto elemType{ |
1172 | mlir::dyn_cast<mlir::FloatType>(cmplx.getElementType())}; |
1173 | width = elemType.getWidth(); |
1174 | }) |
1175 | .template Case<mlir::FloatType>( |
1176 | [&](mlir::FloatType real) { width = real.getWidth(); }); |
1177 | return width; |
1178 | } |
1179 | |
1180 | // Determine if all derived types components are of the same float type with |
1181 | // the same width. Complex(4) is considered 2 floats and complex(8) 2 doubles. |
1182 | bool hasSameFloatAndWidth( |
1183 | fir::RecordType recTy, |
1184 | std::pair<mlir::Type, unsigned> &firstTypeAndWidth) const { |
1185 | for (auto comp : recTy.getTypeList()) { |
1186 | mlir::Type compType{comp.second}; |
1187 | if (mlir::isa<fir::RecordType>(compType)) { |
1188 | auto rc{hasSameFloatAndWidth(mlir::cast<fir::RecordType>(compType), |
1189 | firstTypeAndWidth)}; |
1190 | if (!rc) |
1191 | return false; |
1192 | } else { |
1193 | mlir::Type ty; |
1194 | bool isFloatType{false}; |
1195 | if (mlir::isa<mlir::FloatType, mlir::ComplexType>(compType)) { |
1196 | ty = compType; |
1197 | isFloatType = true; |
1198 | } else if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(compType)) { |
1199 | ty = seqTy.getEleTy(); |
1200 | isFloatType = mlir::isa<mlir::FloatType, mlir::ComplexType>(ty); |
1201 | } |
1202 | |
1203 | if (!isFloatType) { |
1204 | return false; |
1205 | } |
1206 | auto width{getElemWidth(ty)}; |
1207 | if (firstTypeAndWidth.first == nullptr) { |
1208 | firstTypeAndWidth.first = ty; |
1209 | firstTypeAndWidth.second = width; |
1210 | } else if (width != firstTypeAndWidth.second) { |
1211 | return false; |
1212 | } |
1213 | } |
1214 | } |
1215 | return true; |
1216 | } |
1217 | |
1218 | CodeGenSpecifics::Marshalling |
1219 | passOnTheStack(mlir::Location loc, mlir::Type ty, bool isResult) const { |
1220 | CodeGenSpecifics::Marshalling marshal; |
1221 | auto sizeAndAlign{ |
1222 | fir::getTypeSizeAndAlignmentOrCrash(loc, ty, getDataLayout(), kindMap)}; |
1223 | unsigned short align{ |
1224 | std::max(sizeAndAlign.second, static_cast<unsigned short>(8))}; |
1225 | marshal.emplace_back(fir::ReferenceType::get(ty), |
1226 | AT{align, /*byval=*/!isResult, /*sret=*/isResult}); |
1227 | return marshal; |
1228 | } |
1229 | |
1230 | CodeGenSpecifics::Marshalling |
1231 | structType(mlir::Location loc, fir::RecordType recTy, bool isResult) const { |
1232 | CodeGenSpecifics::Marshalling marshal; |
1233 | auto sizeAndAlign{fir::getTypeSizeAndAlignmentOrCrash( |
1234 | loc, recTy, getDataLayout(), kindMap)}; |
1235 | auto recordTypeSize{sizeAndAlign.first}; |
1236 | mlir::Type seqTy; |
1237 | std::pair<mlir::Type, unsigned> firstTyAndWidth{nullptr, 0}; |
1238 | |
1239 | // If there are less than or equal to 8 floats, the structure is flatten as |
1240 | // an array of floats. |
1241 | constexpr uint64_t maxNoOfFloats{8}; |
1242 | |
1243 | // i64 type |
1244 | mlir::Type elemTy{mlir::IntegerType::get(recTy.getContext(), defaultWidth)}; |
1245 | uint64_t nElem{static_cast<uint64_t>( |
1246 | std::ceil(static_cast<float>(recordTypeSize * 8) / defaultWidth))}; |
1247 | |
1248 | // If the derived type components contains are all floats with the same |
1249 | // width, the argument is passed as an array of floats. |
1250 | if (hasSameFloatAndWidth(recTy, firstTyAndWidth)) { |
1251 | uint64_t n{}; |
1252 | auto firstType{firstTyAndWidth.first}; |
1253 | |
1254 | // Type is either float or complex |
1255 | if (auto cmplx = mlir::dyn_cast<mlir::ComplexType>(firstType)) { |
1256 | auto fltType{mlir::dyn_cast<mlir::FloatType>(cmplx.getElementType())}; |
1257 | n = static_cast<uint64_t>(8 * recordTypeSize / fltType.getWidth()); |
1258 | if (n <= maxNoOfFloats) { |
1259 | nElem = n; |
1260 | elemTy = fltType; |
1261 | } |
1262 | } else if (mlir::isa<mlir::FloatType>(firstType)) { |
1263 | auto elemSizeAndAlign{fir::getTypeSizeAndAlignmentOrCrash( |
1264 | loc, firstType, getDataLayout(), kindMap)}; |
1265 | n = static_cast<uint64_t>(recordTypeSize / elemSizeAndAlign.first); |
1266 | if (n <= maxNoOfFloats) { |
1267 | nElem = n; |
1268 | elemTy = firstType; |
1269 | } |
1270 | } |
1271 | // Neither float nor complex |
1272 | assert(n > 0 && "unexpected type" ); |
1273 | } |
1274 | |
1275 | // For function returns, only flattened if there are less than 8 |
1276 | // floats in total. |
1277 | if (isResult && |
1278 | ((mlir::isa<mlir::FloatType>(elemTy) && nElem > maxNoOfFloats) || |
1279 | !mlir::isa<mlir::FloatType>(elemTy))) { |
1280 | return passOnTheStack(loc, recTy, isResult); |
1281 | } |
1282 | |
1283 | seqTy = fir::SequenceType::get(nElem, elemTy); |
1284 | marshal.emplace_back(seqTy, AT{}); |
1285 | return marshal; |
1286 | } |
1287 | |
1288 | CodeGenSpecifics::Marshalling |
1289 | structArgumentType(mlir::Location loc, fir::RecordType recType, |
1290 | const Marshalling &previousArguments) const override { |
1291 | auto sizeAndAlign{fir::getTypeSizeAndAlignmentOrCrash( |
1292 | loc, recType, getDataLayout(), kindMap)}; |
1293 | if (sizeAndAlign.first > 64) { |
1294 | return passOnTheStack(loc, recType, false); |
1295 | } |
1296 | return structType(loc, recType, false); |
1297 | } |
1298 | |
1299 | CodeGenSpecifics::Marshalling |
1300 | structReturnType(mlir::Location loc, fir::RecordType recType) const override { |
1301 | return structType(loc, recType, true); |
1302 | } |
1303 | }; |
1304 | } // namespace |
1305 | |
1306 | //===----------------------------------------------------------------------===// |
1307 | // sparc (sparc 32 bit) target specifics. |
1308 | //===----------------------------------------------------------------------===// |
1309 | |
1310 | namespace { |
1311 | struct TargetSparc : public GenericTarget<TargetSparc> { |
1312 | using GenericTarget::GenericTarget; |
1313 | |
1314 | static constexpr int defaultWidth = 32; |
1315 | |
1316 | CodeGenSpecifics::Marshalling |
1317 | complexArgumentType(mlir::Location, mlir::Type eleTy) const override { |
1318 | assert(fir::isa_real(eleTy)); |
1319 | CodeGenSpecifics::Marshalling marshal; |
1320 | // Use a type that will be translated into LLVM as: |
1321 | // { t, t } struct of 2 eleTy |
1322 | auto structTy = |
1323 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}); |
1324 | marshal.emplace_back(fir::ReferenceType::get(structTy), AT{}); |
1325 | return marshal; |
1326 | } |
1327 | |
1328 | CodeGenSpecifics::Marshalling |
1329 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1330 | assert(fir::isa_real(eleTy)); |
1331 | CodeGenSpecifics::Marshalling marshal; |
1332 | // Use a type that will be translated into LLVM as: |
1333 | // { t, t } struct of 2 eleTy, byval |
1334 | auto structTy = |
1335 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}); |
1336 | marshal.emplace_back(fir::ReferenceType::get(structTy), |
1337 | AT{/*alignment=*/0, /*byval=*/true}); |
1338 | return marshal; |
1339 | } |
1340 | }; |
1341 | } // namespace |
1342 | |
1343 | //===----------------------------------------------------------------------===// |
1344 | // sparcv9 (sparc 64 bit) target specifics. |
1345 | //===----------------------------------------------------------------------===// |
1346 | |
1347 | namespace { |
1348 | struct TargetSparcV9 : public GenericTarget<TargetSparcV9> { |
1349 | using GenericTarget::GenericTarget; |
1350 | |
1351 | static constexpr int defaultWidth = 64; |
1352 | |
1353 | CodeGenSpecifics::Marshalling |
1354 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
1355 | CodeGenSpecifics::Marshalling marshal; |
1356 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
1357 | if (sem == &llvm::APFloat::IEEEsingle() || |
1358 | sem == &llvm::APFloat::IEEEdouble()) { |
1359 | // two distinct float, double arguments |
1360 | marshal.emplace_back(eleTy, AT{}); |
1361 | marshal.emplace_back(eleTy, AT{}); |
1362 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
1363 | // Use a type that will be translated into LLVM as: |
1364 | // { fp128, fp128 } struct of 2 fp128, byval, align 16 |
1365 | marshal.emplace_back( |
1366 | fir::ReferenceType::get(mlir::TupleType::get( |
1367 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
1368 | AT{/*align=*/16, /*byval=*/true}); |
1369 | } else { |
1370 | typeTodo(sem, loc, "argument" ); |
1371 | } |
1372 | return marshal; |
1373 | } |
1374 | |
1375 | CodeGenSpecifics::Marshalling |
1376 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1377 | CodeGenSpecifics::Marshalling marshal; |
1378 | // Use a type that will be translated into LLVM as: |
1379 | // { eleTy, eleTy } struct of 2 eleTy |
1380 | marshal.emplace_back( |
1381 | mlir::TupleType::get(eleTy.getContext(), mlir::TypeRange{eleTy, eleTy}), |
1382 | AT{}); |
1383 | return marshal; |
1384 | } |
1385 | }; |
1386 | } // namespace |
1387 | |
1388 | //===----------------------------------------------------------------------===// |
1389 | // RISCV64 linux target specifics. |
1390 | //===----------------------------------------------------------------------===// |
1391 | |
1392 | namespace { |
1393 | struct TargetRISCV64 : public GenericTarget<TargetRISCV64> { |
1394 | using GenericTarget::GenericTarget; |
1395 | |
1396 | static constexpr int defaultWidth = 64; |
1397 | |
1398 | CodeGenSpecifics::Marshalling |
1399 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
1400 | CodeGenSpecifics::Marshalling marshal; |
1401 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
1402 | if (sem == &llvm::APFloat::IEEEsingle() || |
1403 | sem == &llvm::APFloat::IEEEdouble()) { |
1404 | // Two distinct element type arguments (re, im) |
1405 | marshal.emplace_back(eleTy, AT{}); |
1406 | marshal.emplace_back(eleTy, AT{}); |
1407 | } else { |
1408 | typeTodo(sem, loc, "argument" ); |
1409 | } |
1410 | return marshal; |
1411 | } |
1412 | |
1413 | CodeGenSpecifics::Marshalling |
1414 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1415 | CodeGenSpecifics::Marshalling marshal; |
1416 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
1417 | if (sem == &llvm::APFloat::IEEEsingle() || |
1418 | sem == &llvm::APFloat::IEEEdouble()) { |
1419 | // Use a type that will be translated into LLVM as: |
1420 | // { t, t } struct of 2 eleTy, byVal |
1421 | marshal.emplace_back(mlir::TupleType::get(eleTy.getContext(), |
1422 | mlir::TypeRange{eleTy, eleTy}), |
1423 | AT{/*alignment=*/0, /*byval=*/true}); |
1424 | } else { |
1425 | typeTodo(sem, loc, "return" ); |
1426 | } |
1427 | return marshal; |
1428 | } |
1429 | }; |
1430 | } // namespace |
1431 | |
1432 | //===----------------------------------------------------------------------===// |
1433 | // AMDGPU linux target specifics. |
1434 | //===----------------------------------------------------------------------===// |
1435 | |
1436 | namespace { |
1437 | struct TargetAMDGPU : public GenericTarget<TargetAMDGPU> { |
1438 | using GenericTarget::GenericTarget; |
1439 | |
1440 | // Default size (in bits) of the index type for strings. |
1441 | static constexpr int defaultWidth = 64; |
1442 | |
1443 | CodeGenSpecifics::Marshalling |
1444 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
1445 | CodeGenSpecifics::Marshalling marshal; |
1446 | TODO(loc, "handle complex argument types" ); |
1447 | return marshal; |
1448 | } |
1449 | |
1450 | CodeGenSpecifics::Marshalling |
1451 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1452 | CodeGenSpecifics::Marshalling marshal; |
1453 | TODO(loc, "handle complex return types" ); |
1454 | return marshal; |
1455 | } |
1456 | }; |
1457 | } // namespace |
1458 | |
1459 | //===----------------------------------------------------------------------===// |
1460 | // NVPTX linux target specifics. |
1461 | //===----------------------------------------------------------------------===// |
1462 | |
1463 | namespace { |
1464 | struct TargetNVPTX : public GenericTarget<TargetNVPTX> { |
1465 | using GenericTarget::GenericTarget; |
1466 | |
1467 | // Default size (in bits) of the index type for strings. |
1468 | static constexpr int defaultWidth = 64; |
1469 | |
1470 | CodeGenSpecifics::Marshalling |
1471 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
1472 | CodeGenSpecifics::Marshalling marshal; |
1473 | TODO(loc, "handle complex argument types" ); |
1474 | return marshal; |
1475 | } |
1476 | |
1477 | CodeGenSpecifics::Marshalling |
1478 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1479 | CodeGenSpecifics::Marshalling marshal; |
1480 | TODO(loc, "handle complex return types" ); |
1481 | return marshal; |
1482 | } |
1483 | }; |
1484 | } // namespace |
1485 | |
1486 | //===----------------------------------------------------------------------===// |
1487 | // LoongArch64 linux target specifics. |
1488 | //===----------------------------------------------------------------------===// |
1489 | |
1490 | namespace { |
1491 | struct TargetLoongArch64 : public GenericTarget<TargetLoongArch64> { |
1492 | using GenericTarget::GenericTarget; |
1493 | |
1494 | static constexpr int defaultWidth = 64; |
1495 | static constexpr int GRLen = defaultWidth; /* eight bytes */ |
1496 | static constexpr int GRLenInChar = GRLen / 8; |
1497 | static constexpr int FRLen = defaultWidth; /* eight bytes */ |
1498 | |
1499 | CodeGenSpecifics::Marshalling |
1500 | complexArgumentType(mlir::Location loc, mlir::Type eleTy) const override { |
1501 | CodeGenSpecifics::Marshalling marshal; |
1502 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
1503 | if (sem == &llvm::APFloat::IEEEsingle() || |
1504 | sem == &llvm::APFloat::IEEEdouble()) { |
1505 | // Two distinct element type arguments (re, im) |
1506 | marshal.emplace_back(eleTy, AT{}); |
1507 | marshal.emplace_back(eleTy, AT{}); |
1508 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
1509 | // Use a type that will be translated into LLVM as: |
1510 | // { fp128, fp128 } struct of 2 fp128, byval |
1511 | marshal.emplace_back( |
1512 | fir::ReferenceType::get(mlir::TupleType::get( |
1513 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
1514 | AT{/*align=*/16, /*byval=*/true}); |
1515 | } else { |
1516 | typeTodo(sem, loc, "argument" ); |
1517 | } |
1518 | return marshal; |
1519 | } |
1520 | |
1521 | CodeGenSpecifics::Marshalling |
1522 | complexReturnType(mlir::Location loc, mlir::Type eleTy) const override { |
1523 | CodeGenSpecifics::Marshalling marshal; |
1524 | const auto *sem = &floatToSemantics(kindMap, eleTy); |
1525 | if (sem == &llvm::APFloat::IEEEsingle() || |
1526 | sem == &llvm::APFloat::IEEEdouble()) { |
1527 | // Use a type that will be translated into LLVM as: |
1528 | // { t, t } struct of 2 eleTy, byVal |
1529 | marshal.emplace_back(mlir::TupleType::get(eleTy.getContext(), |
1530 | mlir::TypeRange{eleTy, eleTy}), |
1531 | AT{/*alignment=*/0, /*byval=*/true}); |
1532 | } else if (sem == &llvm::APFloat::IEEEquad()) { |
1533 | // Use a type that will be translated into LLVM as: |
1534 | // { fp128, fp128 } struct of 2 fp128, sret, align 16 |
1535 | marshal.emplace_back( |
1536 | fir::ReferenceType::get(mlir::TupleType::get( |
1537 | eleTy.getContext(), mlir::TypeRange{eleTy, eleTy})), |
1538 | AT{/*align=*/16, /*byval=*/false, /*sret=*/true}); |
1539 | } else { |
1540 | typeTodo(sem, loc, "return" ); |
1541 | } |
1542 | return marshal; |
1543 | } |
1544 | |
1545 | CodeGenSpecifics::Marshalling |
1546 | integerArgumentType(mlir::Location loc, |
1547 | mlir::IntegerType argTy) const override { |
1548 | if (argTy.getWidth() == 32) { |
1549 | // LA64 LP64D ABI requires unsigned 32 bit integers to be sign extended. |
1550 | // Therefore, Flang also follows it if a function needs to be |
1551 | // interoperable with C. |
1552 | // |
1553 | // Currently, it only adds `signext` attribute to the dummy arguments and |
1554 | // return values in the function signatures, but it does not add the |
1555 | // corresponding attribute to the actual arguments and return values in |
1556 | // `fir.call` instruction. Thanks to LLVM's integration of all these |
1557 | // attributes, the modification is still effective. |
1558 | CodeGenSpecifics::Marshalling marshal; |
1559 | AT::IntegerExtension intExt = AT::IntegerExtension::Sign; |
1560 | marshal.emplace_back(argTy, AT{/*alignment=*/0, /*byval=*/false, |
1561 | /*sret=*/false, /*append=*/false, |
1562 | /*intExt=*/intExt}); |
1563 | return marshal; |
1564 | } |
1565 | |
1566 | return GenericTarget::integerArgumentType(loc, argTy); |
1567 | } |
1568 | |
1569 | /// Flatten non-basic types, resulting in an array of types containing only |
1570 | /// `IntegerType` and `FloatType`. |
1571 | llvm::SmallVector<mlir::Type> flattenTypeList(mlir::Location loc, |
1572 | const mlir::Type type) const { |
1573 | llvm::SmallVector<mlir::Type> flatTypes; |
1574 | |
1575 | llvm::TypeSwitch<mlir::Type>(type) |
1576 | .template Case<mlir::IntegerType>([&](mlir::IntegerType intTy) { |
1577 | if (intTy.getWidth() != 0) |
1578 | flatTypes.push_back(intTy); |
1579 | }) |
1580 | .template Case<mlir::FloatType>([&](mlir::FloatType floatTy) { |
1581 | if (floatTy.getWidth() != 0) |
1582 | flatTypes.push_back(floatTy); |
1583 | }) |
1584 | .template Case<mlir::ComplexType>([&](mlir::ComplexType cmplx) { |
1585 | const auto *sem = &floatToSemantics(kindMap, cmplx.getElementType()); |
1586 | if (sem == &llvm::APFloat::IEEEsingle() || |
1587 | sem == &llvm::APFloat::IEEEdouble() || |
1588 | sem == &llvm::APFloat::IEEEquad()) |
1589 | std::fill_n(std::back_inserter(flatTypes), 2, |
1590 | cmplx.getElementType()); |
1591 | else |
1592 | TODO(loc, "unsupported complex type(not IEEEsingle, IEEEdouble, " |
1593 | "IEEEquad) as a structure component for BIND(C), " |
1594 | "VALUE derived type argument and type return" ); |
1595 | }) |
1596 | .template Case<fir::LogicalType>([&](fir::LogicalType logicalTy) { |
1597 | const unsigned width = |
1598 | kindMap.getLogicalBitsize(logicalTy.getFKind()); |
1599 | if (width != 0) |
1600 | flatTypes.push_back( |
1601 | mlir::IntegerType::get(type.getContext(), width)); |
1602 | }) |
1603 | .template Case<fir::CharacterType>([&](fir::CharacterType charTy) { |
1604 | assert(kindMap.getCharacterBitsize(charTy.getFKind()) <= 8 && |
1605 | "the bit size of characterType as an interoperable type must " |
1606 | "not exceed 8" ); |
1607 | for (unsigned i = 0; i < charTy.getLen(); ++i) |
1608 | flatTypes.push_back(mlir::IntegerType::get(type.getContext(), 8)); |
1609 | }) |
1610 | .template Case<fir::SequenceType>([&](fir::SequenceType seqTy) { |
1611 | if (!seqTy.hasDynamicExtents()) { |
1612 | const std::uint64_t numOfEle = seqTy.getConstantArraySize(); |
1613 | mlir::Type eleTy = seqTy.getEleTy(); |
1614 | if (!mlir::isa<mlir::IntegerType, mlir::FloatType>(eleTy)) { |
1615 | llvm::SmallVector<mlir::Type> subTypeList = |
1616 | flattenTypeList(loc, eleTy); |
1617 | if (subTypeList.size() != 0) |
1618 | for (std::uint64_t i = 0; i < numOfEle; ++i) |
1619 | llvm::copy(subTypeList, std::back_inserter(flatTypes)); |
1620 | } else { |
1621 | std::fill_n(std::back_inserter(flatTypes), numOfEle, eleTy); |
1622 | } |
1623 | } else |
1624 | TODO(loc, "unsupported dynamic extent sequence type as a structure " |
1625 | "component for BIND(C), " |
1626 | "VALUE derived type argument and type return" ); |
1627 | }) |
1628 | .template Case<fir::RecordType>([&](fir::RecordType recTy) { |
1629 | for (auto &component : recTy.getTypeList()) { |
1630 | mlir::Type eleTy = component.second; |
1631 | llvm::SmallVector<mlir::Type> subTypeList = |
1632 | flattenTypeList(loc, eleTy); |
1633 | if (subTypeList.size() != 0) |
1634 | llvm::copy(subTypeList, std::back_inserter(flatTypes)); |
1635 | } |
1636 | }) |
1637 | .template Case<fir::VectorType>([&](fir::VectorType vecTy) { |
1638 | auto sizeAndAlign = fir::getTypeSizeAndAlignmentOrCrash( |
1639 | loc, vecTy, getDataLayout(), kindMap); |
1640 | if (sizeAndAlign.first == 2 * GRLenInChar) |
1641 | flatTypes.push_back( |
1642 | mlir::IntegerType::get(type.getContext(), 2 * GRLen)); |
1643 | else |
1644 | TODO(loc, "unsupported vector width(must be 128 bits)" ); |
1645 | }) |
1646 | .Default([&](mlir::Type ty) { |
1647 | if (fir::conformsWithPassByRef(ty)) |
1648 | flatTypes.push_back( |
1649 | mlir::IntegerType::get(type.getContext(), GRLen)); |
1650 | else |
1651 | TODO(loc, "unsupported component type for BIND(C), VALUE derived " |
1652 | "type argument and type return" ); |
1653 | }); |
1654 | |
1655 | return flatTypes; |
1656 | } |
1657 | |
1658 | /// Determine if a struct is eligible to be passed in FARs (and GARs) (i.e., |
1659 | /// when flattened it contains a single fp value, fp+fp, or int+fp of |
1660 | /// appropriate size). |
1661 | bool detectFARsEligibleStruct(mlir::Location loc, fir::RecordType recTy, |
1662 | mlir::Type &field1Ty, |
1663 | mlir::Type &field2Ty) const { |
1664 | field1Ty = field2Ty = nullptr; |
1665 | llvm::SmallVector<mlir::Type> flatTypes = flattenTypeList(loc, recTy); |
1666 | size_t flatSize = flatTypes.size(); |
1667 | |
1668 | // Cannot be eligible if the number of flattened types is equal to 0 or |
1669 | // greater than 2. |
1670 | if (flatSize == 0 || flatSize > 2) |
1671 | return false; |
1672 | |
1673 | bool isFirstAvaliableFloat = false; |
1674 | |
1675 | assert((mlir::isa<mlir::IntegerType, mlir::FloatType>(flatTypes[0])) && |
1676 | "Type must be integerType or floatType after flattening" ); |
1677 | if (auto floatTy = mlir::dyn_cast<mlir::FloatType>(flatTypes[0])) { |
1678 | const unsigned Size = floatTy.getWidth(); |
1679 | // Can't be eligible if larger than the FP registers. Half precision isn't |
1680 | // currently supported on LoongArch and the ABI hasn't been confirmed, so |
1681 | // default to the integer ABI in that case. |
1682 | if (Size > FRLen || Size < 32) |
1683 | return false; |
1684 | isFirstAvaliableFloat = true; |
1685 | field1Ty = floatTy; |
1686 | } else if (auto intTy = mlir::dyn_cast<mlir::IntegerType>(flatTypes[0])) { |
1687 | if (intTy.getWidth() > GRLen) |
1688 | return false; |
1689 | field1Ty = intTy; |
1690 | } |
1691 | |
1692 | // flatTypes has two elements |
1693 | if (flatSize == 2) { |
1694 | assert((mlir::isa<mlir::IntegerType, mlir::FloatType>(flatTypes[1])) && |
1695 | "Type must be integerType or floatType after flattening" ); |
1696 | if (auto floatTy = mlir::dyn_cast<mlir::FloatType>(flatTypes[1])) { |
1697 | const unsigned Size = floatTy.getWidth(); |
1698 | if (Size > FRLen || Size < 32) |
1699 | return false; |
1700 | field2Ty = floatTy; |
1701 | return true; |
1702 | } else if (auto intTy = mlir::dyn_cast<mlir::IntegerType>(flatTypes[1])) { |
1703 | // Can't be eligible if an integer type was already found (int+int pairs |
1704 | // are not eligible). |
1705 | if (!isFirstAvaliableFloat) |
1706 | return false; |
1707 | if (intTy.getWidth() > GRLen) |
1708 | return false; |
1709 | field2Ty = intTy; |
1710 | return true; |
1711 | } |
1712 | } |
1713 | |
1714 | // return isFirstAvaliableFloat if flatTypes only has one element |
1715 | return isFirstAvaliableFloat; |
1716 | } |
1717 | |
1718 | bool checkTypeHasEnoughRegs(mlir::Location loc, int &GARsLeft, int &FARsLeft, |
1719 | const mlir::Type type) const { |
1720 | if (!type) |
1721 | return true; |
1722 | |
1723 | llvm::TypeSwitch<mlir::Type>(type) |
1724 | .template Case<mlir::IntegerType>([&](mlir::IntegerType intTy) { |
1725 | const unsigned width = intTy.getWidth(); |
1726 | if (width > 128) |
1727 | TODO(loc, |
1728 | "integerType with width exceeding 128 bits is unsupported" ); |
1729 | if (width == 0) |
1730 | return; |
1731 | if (width <= GRLen) |
1732 | --GARsLeft; |
1733 | else if (width <= 2 * GRLen) |
1734 | GARsLeft = GARsLeft - 2; |
1735 | }) |
1736 | .template Case<mlir::FloatType>([&](mlir::FloatType floatTy) { |
1737 | const unsigned width = floatTy.getWidth(); |
1738 | if (width > 128) |
1739 | TODO(loc, "floatType with width exceeding 128 bits is unsupported" ); |
1740 | if (width == 0) |
1741 | return; |
1742 | if (width == 32 || width == 64) |
1743 | --FARsLeft; |
1744 | else if (width <= GRLen) |
1745 | --GARsLeft; |
1746 | else if (width <= 2 * GRLen) |
1747 | GARsLeft = GARsLeft - 2; |
1748 | }) |
1749 | .Default([&](mlir::Type ty) { |
1750 | if (fir::conformsWithPassByRef(ty)) |
1751 | --GARsLeft; // Pointers. |
1752 | else |
1753 | TODO(loc, "unsupported component type for BIND(C), VALUE derived " |
1754 | "type argument and type return" ); |
1755 | }); |
1756 | |
1757 | return GARsLeft >= 0 && FARsLeft >= 0; |
1758 | } |
1759 | |
1760 | bool hasEnoughRegisters(mlir::Location loc, int GARsLeft, int FARsLeft, |
1761 | const Marshalling &previousArguments, |
1762 | const mlir::Type &field1Ty, |
1763 | const mlir::Type &field2Ty) const { |
1764 | for (auto &typeAndAttr : previousArguments) { |
1765 | const auto &attr = std::get<Attributes>(typeAndAttr); |
1766 | if (attr.isByVal()) { |
1767 | // Previous argument passed on the stack, and its address is passed in |
1768 | // GAR. |
1769 | --GARsLeft; |
1770 | continue; |
1771 | } |
1772 | |
1773 | // Previous aggregate arguments were marshalled into simpler arguments. |
1774 | const auto &type = std::get<mlir::Type>(typeAndAttr); |
1775 | llvm::SmallVector<mlir::Type> flatTypes = flattenTypeList(loc, type); |
1776 | |
1777 | for (auto &flatTy : flatTypes) { |
1778 | if (!checkTypeHasEnoughRegs(loc, GARsLeft, FARsLeft, flatTy)) |
1779 | return false; |
1780 | } |
1781 | } |
1782 | |
1783 | if (!checkTypeHasEnoughRegs(loc, GARsLeft, FARsLeft, field1Ty)) |
1784 | return false; |
1785 | if (!checkTypeHasEnoughRegs(loc, GARsLeft, FARsLeft, field2Ty)) |
1786 | return false; |
1787 | return true; |
1788 | } |
1789 | |
1790 | /// LoongArch64 subroutine calling sequence ABI in: |
1791 | /// https://github.com/loongson/la-abi-specs/blob/release/lapcs.adoc#subroutine-calling-sequence |
1792 | CodeGenSpecifics::Marshalling |
1793 | classifyStruct(mlir::Location loc, fir::RecordType recTy, int GARsLeft, |
1794 | int FARsLeft, bool isResult, |
1795 | const Marshalling &previousArguments) const { |
1796 | CodeGenSpecifics::Marshalling marshal; |
1797 | |
1798 | auto [recSize, recAlign] = fir::getTypeSizeAndAlignmentOrCrash( |
1799 | loc, recTy, getDataLayout(), kindMap); |
1800 | mlir::MLIRContext *context = recTy.getContext(); |
1801 | |
1802 | if (recSize == 0) { |
1803 | TODO(loc, "unsupported empty struct type for BIND(C), " |
1804 | "VALUE derived type argument and type return" ); |
1805 | } |
1806 | |
1807 | if (recSize > 2 * GRLenInChar) { |
1808 | marshal.emplace_back( |
1809 | fir::ReferenceType::get(recTy), |
1810 | AT{recAlign, /*byval=*/!isResult, /*sret=*/isResult}); |
1811 | return marshal; |
1812 | } |
1813 | |
1814 | // Pass by FARs(and GARs) |
1815 | mlir::Type field1Ty = nullptr, field2Ty = nullptr; |
1816 | if (detectFARsEligibleStruct(loc, recTy, field1Ty, field2Ty) && |
1817 | hasEnoughRegisters(loc, GARsLeft, FARsLeft, previousArguments, field1Ty, |
1818 | field2Ty)) { |
1819 | if (!isResult) { |
1820 | if (field1Ty) |
1821 | marshal.emplace_back(field1Ty, AT{}); |
1822 | if (field2Ty) |
1823 | marshal.emplace_back(field2Ty, AT{}); |
1824 | } else { |
1825 | // field1Ty is always preferred over field2Ty for assignment, so there |
1826 | // will never be a case where field1Ty == nullptr and field2Ty != |
1827 | // nullptr. |
1828 | if (field1Ty && !field2Ty) |
1829 | marshal.emplace_back(field1Ty, AT{}); |
1830 | else if (field1Ty && field2Ty) |
1831 | marshal.emplace_back( |
1832 | mlir::TupleType::get(context, |
1833 | mlir::TypeRange{field1Ty, field2Ty}), |
1834 | AT{/*alignment=*/0, /*byval=*/true}); |
1835 | } |
1836 | return marshal; |
1837 | } |
1838 | |
1839 | if (recSize <= GRLenInChar) { |
1840 | marshal.emplace_back(mlir::IntegerType::get(context, GRLen), AT{}); |
1841 | return marshal; |
1842 | } |
1843 | |
1844 | if (recAlign == 2 * GRLenInChar) { |
1845 | marshal.emplace_back(mlir::IntegerType::get(context, 2 * GRLen), AT{}); |
1846 | return marshal; |
1847 | } |
1848 | |
1849 | // recSize > GRLenInChar && recSize <= 2 * GRLenInChar |
1850 | marshal.emplace_back( |
1851 | fir::SequenceType::get({2}, mlir::IntegerType::get(context, GRLen)), |
1852 | AT{}); |
1853 | return marshal; |
1854 | } |
1855 | |
1856 | /// Marshal a derived type passed by value like a C struct. |
1857 | CodeGenSpecifics::Marshalling |
1858 | structArgumentType(mlir::Location loc, fir::RecordType recTy, |
1859 | const Marshalling &previousArguments) const override { |
1860 | int GARsLeft = 8; |
1861 | int FARsLeft = FRLen ? 8 : 0; |
1862 | |
1863 | return classifyStruct(loc, recTy, GARsLeft, FARsLeft, /*isResult=*/false, |
1864 | previousArguments); |
1865 | } |
1866 | |
1867 | CodeGenSpecifics::Marshalling |
1868 | structReturnType(mlir::Location loc, fir::RecordType recTy) const override { |
1869 | // The rules for return and argument types are the same. |
1870 | int GARsLeft = 2; |
1871 | int FARsLeft = FRLen ? 2 : 0; |
1872 | return classifyStruct(loc, recTy, GARsLeft, FARsLeft, /*isResult=*/true, |
1873 | {}); |
1874 | } |
1875 | }; |
1876 | } // namespace |
1877 | |
1878 | // Instantiate the overloaded target instance based on the triple value. |
1879 | // TODO: Add other targets to this file as needed. |
1880 | std::unique_ptr<fir::CodeGenSpecifics> |
1881 | fir::CodeGenSpecifics::get(mlir::MLIRContext *ctx, llvm::Triple &&trp, |
1882 | KindMapping &&kindMap, llvm::StringRef targetCPU, |
1883 | mlir::LLVM::TargetFeaturesAttr targetFeatures, |
1884 | const mlir::DataLayout &dl) { |
1885 | switch (trp.getArch()) { |
1886 | default: |
1887 | break; |
1888 | case llvm::Triple::ArchType::x86: |
1889 | if (trp.isOSWindows()) |
1890 | return std::make_unique<TargetI386Win>(ctx, std::move(trp), |
1891 | std::move(kindMap), targetCPU, |
1892 | targetFeatures, dl); |
1893 | else |
1894 | return std::make_unique<TargetI386>(ctx, std::move(trp), |
1895 | std::move(kindMap), targetCPU, |
1896 | targetFeatures, dl); |
1897 | case llvm::Triple::ArchType::x86_64: |
1898 | if (trp.isOSWindows()) |
1899 | return std::make_unique<TargetX86_64Win>(ctx, std::move(trp), |
1900 | std::move(kindMap), targetCPU, |
1901 | targetFeatures, dl); |
1902 | else |
1903 | return std::make_unique<TargetX86_64>(ctx, std::move(trp), |
1904 | std::move(kindMap), targetCPU, |
1905 | targetFeatures, dl); |
1906 | case llvm::Triple::ArchType::aarch64: |
1907 | return std::make_unique<TargetAArch64>( |
1908 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1909 | case llvm::Triple::ArchType::ppc: |
1910 | return std::make_unique<TargetPPC>(ctx, std::move(trp), std::move(kindMap), |
1911 | targetCPU, targetFeatures, dl); |
1912 | case llvm::Triple::ArchType::ppc64: |
1913 | return std::make_unique<TargetPPC64>( |
1914 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1915 | case llvm::Triple::ArchType::ppc64le: |
1916 | return std::make_unique<TargetPPC64le>( |
1917 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1918 | case llvm::Triple::ArchType::sparc: |
1919 | return std::make_unique<TargetSparc>( |
1920 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1921 | case llvm::Triple::ArchType::sparcv9: |
1922 | return std::make_unique<TargetSparcV9>( |
1923 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1924 | case llvm::Triple::ArchType::riscv64: |
1925 | return std::make_unique<TargetRISCV64>( |
1926 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1927 | case llvm::Triple::ArchType::amdgcn: |
1928 | return std::make_unique<TargetAMDGPU>( |
1929 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1930 | case llvm::Triple::ArchType::nvptx64: |
1931 | return std::make_unique<TargetNVPTX>( |
1932 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1933 | case llvm::Triple::ArchType::loongarch64: |
1934 | return std::make_unique<TargetLoongArch64>( |
1935 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1936 | } |
1937 | TODO(mlir::UnknownLoc::get(ctx), "target not implemented" ); |
1938 | } |
1939 | |
1940 | std::unique_ptr<fir::CodeGenSpecifics> fir::CodeGenSpecifics::get( |
1941 | mlir::MLIRContext *ctx, llvm::Triple &&trp, KindMapping &&kindMap, |
1942 | llvm::StringRef targetCPU, mlir::LLVM::TargetFeaturesAttr targetFeatures, |
1943 | const mlir::DataLayout &dl, llvm::StringRef tuneCPU) { |
1944 | std::unique_ptr<fir::CodeGenSpecifics> CGS = fir::CodeGenSpecifics::get( |
1945 | ctx, std::move(trp), std::move(kindMap), targetCPU, targetFeatures, dl); |
1946 | |
1947 | CGS->tuneCPU = tuneCPU; |
1948 | return CGS; |
1949 | } |
1950 | |