1 | //===-- CodeGen.cpp -- bridge to lower to LLVM ----------------------------===// |
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/CodeGen.h" |
14 | |
15 | #include "CGOps.h" |
16 | #include "flang/Optimizer/CodeGen/CodeGenOpenMP.h" |
17 | #include "flang/Optimizer/CodeGen/FIROpPatterns.h" |
18 | #include "flang/Optimizer/CodeGen/TypeConverter.h" |
19 | #include "flang/Optimizer/Dialect/FIRAttr.h" |
20 | #include "flang/Optimizer/Dialect/FIROps.h" |
21 | #include "flang/Optimizer/Dialect/FIRType.h" |
22 | #include "flang/Optimizer/Support/DataLayout.h" |
23 | #include "flang/Optimizer/Support/InternalNames.h" |
24 | #include "flang/Optimizer/Support/TypeCode.h" |
25 | #include "flang/Optimizer/Support/Utils.h" |
26 | #include "flang/Semantics/runtime-type-info.h" |
27 | #include "mlir/Conversion/ArithCommon/AttrToLLVMConverter.h" |
28 | #include "mlir/Conversion/ArithToLLVM/ArithToLLVM.h" |
29 | #include "mlir/Conversion/ComplexToLLVM/ComplexToLLVM.h" |
30 | #include "mlir/Conversion/ComplexToStandard/ComplexToStandard.h" |
31 | #include "mlir/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.h" |
32 | #include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVM.h" |
33 | #include "mlir/Conversion/LLVMCommon/Pattern.h" |
34 | #include "mlir/Conversion/MathToFuncs/MathToFuncs.h" |
35 | #include "mlir/Conversion/MathToLLVM/MathToLLVM.h" |
36 | #include "mlir/Conversion/MathToLibm/MathToLibm.h" |
37 | #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h" |
38 | #include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h" |
39 | #include "mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h" |
40 | #include "mlir/Dialect/Arith/IR/Arith.h" |
41 | #include "mlir/Dialect/DLTI/DLTI.h" |
42 | #include "mlir/Dialect/LLVMIR/LLVMAttrs.h" |
43 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" |
44 | #include "mlir/Dialect/LLVMIR/Transforms/AddComdats.h" |
45 | #include "mlir/Dialect/OpenACC/OpenACC.h" |
46 | #include "mlir/Dialect/OpenMP/OpenMPDialect.h" |
47 | #include "mlir/IR/BuiltinTypes.h" |
48 | #include "mlir/IR/Matchers.h" |
49 | #include "mlir/Pass/Pass.h" |
50 | #include "mlir/Pass/PassManager.h" |
51 | #include "mlir/Target/LLVMIR/Import.h" |
52 | #include "mlir/Target/LLVMIR/ModuleTranslation.h" |
53 | #include "llvm/ADT/ArrayRef.h" |
54 | #include "llvm/ADT/TypeSwitch.h" |
55 | |
56 | namespace fir { |
57 | #define GEN_PASS_DEF_FIRTOLLVMLOWERING |
58 | #include "flang/Optimizer/CodeGen/CGPasses.h.inc" |
59 | } // namespace fir |
60 | |
61 | #define DEBUG_TYPE "flang-codegen" |
62 | |
63 | // TODO: This should really be recovered from the specified target. |
64 | static constexpr unsigned defaultAlign = 8; |
65 | |
66 | /// `fir.box` attribute values as defined for CFI_attribute_t in |
67 | /// flang/ISO_Fortran_binding.h. |
68 | static constexpr unsigned kAttrPointer = CFI_attribute_pointer; |
69 | static constexpr unsigned kAttrAllocatable = CFI_attribute_allocatable; |
70 | |
71 | static inline mlir::Type getLlvmPtrType(mlir::MLIRContext *context, |
72 | unsigned addressSpace = 0) { |
73 | return mlir::LLVM::LLVMPointerType::get(context, addressSpace); |
74 | } |
75 | |
76 | static inline mlir::Type getI8Type(mlir::MLIRContext *context) { |
77 | return mlir::IntegerType::get(context, 8); |
78 | } |
79 | |
80 | static mlir::LLVM::ConstantOp |
81 | genConstantIndex(mlir::Location loc, mlir::Type ity, |
82 | mlir::ConversionPatternRewriter &rewriter, |
83 | std::int64_t offset) { |
84 | auto cattr = rewriter.getI64IntegerAttr(offset); |
85 | return rewriter.create<mlir::LLVM::ConstantOp>(loc, ity, cattr); |
86 | } |
87 | |
88 | static mlir::Block *createBlock(mlir::ConversionPatternRewriter &rewriter, |
89 | mlir::Block *insertBefore) { |
90 | assert(insertBefore && "expected valid insertion block" ); |
91 | return rewriter.createBlock(insertBefore->getParent(), |
92 | mlir::Region::iterator(insertBefore)); |
93 | } |
94 | |
95 | /// Extract constant from a value that must be the result of one of the |
96 | /// ConstantOp operations. |
97 | static int64_t getConstantIntValue(mlir::Value val) { |
98 | if (auto constVal = fir::getIntIfConstant(val)) |
99 | return *constVal; |
100 | fir::emitFatalError(val.getLoc(), "must be a constant" ); |
101 | } |
102 | |
103 | static unsigned getTypeDescFieldId(mlir::Type ty) { |
104 | auto isArray = fir::dyn_cast_ptrOrBoxEleTy(ty).isa<fir::SequenceType>(); |
105 | return isArray ? kOptTypePtrPosInBox : kDimsPosInBox; |
106 | } |
107 | static unsigned getLenParamFieldId(mlir::Type ty) { |
108 | return getTypeDescFieldId(ty) + 1; |
109 | } |
110 | |
111 | namespace { |
112 | /// Lower `fir.address_of` operation to `llvm.address_of` operation. |
113 | struct AddrOfOpConversion : public fir::FIROpConversion<fir::AddrOfOp> { |
114 | using FIROpConversion::FIROpConversion; |
115 | |
116 | mlir::LogicalResult |
117 | matchAndRewrite(fir::AddrOfOp addr, OpAdaptor adaptor, |
118 | mlir::ConversionPatternRewriter &rewriter) const override { |
119 | auto ty = convertType(addr.getType()); |
120 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>( |
121 | addr, ty, addr.getSymbol().getRootReference().getValue()); |
122 | return mlir::success(); |
123 | } |
124 | }; |
125 | } // namespace |
126 | |
127 | /// Lookup the function to compute the memory size of this parametric derived |
128 | /// type. The size of the object may depend on the LEN type parameters of the |
129 | /// derived type. |
130 | static mlir::LLVM::LLVMFuncOp |
131 | getDependentTypeMemSizeFn(fir::RecordType recTy, fir::AllocaOp op, |
132 | mlir::ConversionPatternRewriter &rewriter) { |
133 | auto module = op->getParentOfType<mlir::ModuleOp>(); |
134 | std::string name = recTy.getName().str() + "P.mem.size" ; |
135 | if (auto memSizeFunc = module.lookupSymbol<mlir::LLVM::LLVMFuncOp>(name)) |
136 | return memSizeFunc; |
137 | TODO(op.getLoc(), "did not find allocation function" ); |
138 | } |
139 | |
140 | // Compute the alloc scale size (constant factors encoded in the array type). |
141 | // We do this for arrays without a constant interior or arrays of character with |
142 | // dynamic length arrays, since those are the only ones that get decayed to a |
143 | // pointer to the element type. |
144 | template <typename OP> |
145 | static mlir::Value |
146 | genAllocationScaleSize(OP op, mlir::Type ity, |
147 | mlir::ConversionPatternRewriter &rewriter) { |
148 | mlir::Location loc = op.getLoc(); |
149 | mlir::Type dataTy = op.getInType(); |
150 | auto seqTy = dataTy.dyn_cast<fir::SequenceType>(); |
151 | fir::SequenceType::Extent constSize = 1; |
152 | if (seqTy) { |
153 | int constRows = seqTy.getConstantRows(); |
154 | const fir::SequenceType::ShapeRef &shape = seqTy.getShape(); |
155 | if (constRows != static_cast<int>(shape.size())) { |
156 | for (auto extent : shape) { |
157 | if (constRows-- > 0) |
158 | continue; |
159 | if (extent != fir::SequenceType::getUnknownExtent()) |
160 | constSize *= extent; |
161 | } |
162 | } |
163 | } |
164 | |
165 | if (constSize != 1) { |
166 | mlir::Value constVal{ |
167 | genConstantIndex(loc, ity, rewriter, constSize).getResult()}; |
168 | return constVal; |
169 | } |
170 | return nullptr; |
171 | } |
172 | |
173 | namespace { |
174 | /// convert to LLVM IR dialect `alloca` |
175 | struct AllocaOpConversion : public fir::FIROpConversion<fir::AllocaOp> { |
176 | using FIROpConversion::FIROpConversion; |
177 | |
178 | mlir::LogicalResult |
179 | matchAndRewrite(fir::AllocaOp alloc, OpAdaptor adaptor, |
180 | mlir::ConversionPatternRewriter &rewriter) const override { |
181 | mlir::ValueRange operands = adaptor.getOperands(); |
182 | auto loc = alloc.getLoc(); |
183 | mlir::Type ity = lowerTy().indexType(); |
184 | unsigned i = 0; |
185 | mlir::Value size = genConstantIndex(loc, ity, rewriter, 1).getResult(); |
186 | mlir::Type firObjType = fir::unwrapRefType(alloc.getType()); |
187 | mlir::Type llvmObjectType = convertObjectType(firObjType); |
188 | if (alloc.hasLenParams()) { |
189 | unsigned end = alloc.numLenParams(); |
190 | llvm::SmallVector<mlir::Value> lenParams; |
191 | for (; i < end; ++i) |
192 | lenParams.push_back(operands[i]); |
193 | mlir::Type scalarType = fir::unwrapSequenceType(alloc.getInType()); |
194 | if (auto chrTy = scalarType.dyn_cast<fir::CharacterType>()) { |
195 | fir::CharacterType rawCharTy = fir::CharacterType::getUnknownLen( |
196 | chrTy.getContext(), chrTy.getFKind()); |
197 | llvmObjectType = convertType(rawCharTy); |
198 | assert(end == 1); |
199 | size = integerCast(loc, rewriter, ity, lenParams[0]); |
200 | } else if (auto recTy = scalarType.dyn_cast<fir::RecordType>()) { |
201 | mlir::LLVM::LLVMFuncOp memSizeFn = |
202 | getDependentTypeMemSizeFn(recTy, alloc, rewriter); |
203 | if (!memSizeFn) |
204 | emitError(loc, "did not find allocation function" ); |
205 | mlir::NamedAttribute attr = rewriter.getNamedAttr( |
206 | "callee" , mlir::SymbolRefAttr::get(memSizeFn)); |
207 | auto call = rewriter.create<mlir::LLVM::CallOp>( |
208 | loc, ity, lenParams, llvm::ArrayRef<mlir::NamedAttribute>{attr}); |
209 | size = call.getResult(); |
210 | llvmObjectType = ::getI8Type(alloc.getContext()); |
211 | } else { |
212 | return emitError(loc, "unexpected type " ) |
213 | << scalarType << " with type parameters" ; |
214 | } |
215 | } |
216 | if (auto scaleSize = genAllocationScaleSize(alloc, ity, rewriter)) |
217 | size = rewriter.create<mlir::LLVM::MulOp>(loc, ity, size, scaleSize); |
218 | if (alloc.hasShapeOperands()) { |
219 | unsigned end = operands.size(); |
220 | for (; i < end; ++i) |
221 | size = rewriter.create<mlir::LLVM::MulOp>( |
222 | loc, ity, size, integerCast(loc, rewriter, ity, operands[i])); |
223 | } |
224 | |
225 | unsigned allocaAs = getAllocaAddressSpace(rewriter); |
226 | unsigned programAs = getProgramAddressSpace(rewriter); |
227 | |
228 | // NOTE: we used to pass alloc->getAttrs() in the builder for non opaque |
229 | // pointers! Only propagate pinned and bindc_name to help debugging, but |
230 | // this should have no functional purpose (and passing the operand segment |
231 | // attribute like before is certainly bad). |
232 | auto llvmAlloc = rewriter.create<mlir::LLVM::AllocaOp>( |
233 | loc, ::getLlvmPtrType(alloc.getContext(), allocaAs), llvmObjectType, |
234 | size); |
235 | if (alloc.getPinned()) |
236 | llvmAlloc->setDiscardableAttr(alloc.getPinnedAttrName(), |
237 | alloc.getPinnedAttr()); |
238 | if (alloc.getBindcName()) |
239 | llvmAlloc->setDiscardableAttr(alloc.getBindcNameAttrName(), |
240 | alloc.getBindcNameAttr()); |
241 | if (allocaAs == programAs) { |
242 | rewriter.replaceOp(alloc, llvmAlloc); |
243 | } else { |
244 | // if our allocation address space, is not the same as the program address |
245 | // space, then we must emit a cast to the program address space before |
246 | // use. An example case would be on AMDGPU, where the allocation address |
247 | // space is the numeric value 5 (private), and the program address space |
248 | // is 0 (generic). |
249 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddrSpaceCastOp>( |
250 | alloc, ::getLlvmPtrType(alloc.getContext(), programAs), llvmAlloc); |
251 | } |
252 | return mlir::success(); |
253 | } |
254 | }; |
255 | } // namespace |
256 | |
257 | namespace { |
258 | /// Lower `fir.box_addr` to the sequence of operations to extract the first |
259 | /// element of the box. |
260 | struct BoxAddrOpConversion : public fir::FIROpConversion<fir::BoxAddrOp> { |
261 | using FIROpConversion::FIROpConversion; |
262 | |
263 | mlir::LogicalResult |
264 | matchAndRewrite(fir::BoxAddrOp boxaddr, OpAdaptor adaptor, |
265 | mlir::ConversionPatternRewriter &rewriter) const override { |
266 | mlir::Value a = adaptor.getOperands()[0]; |
267 | auto loc = boxaddr.getLoc(); |
268 | if (auto argty = boxaddr.getVal().getType().dyn_cast<fir::BaseBoxType>()) { |
269 | TypePair boxTyPair = getBoxTypePair(argty); |
270 | rewriter.replaceOp(boxaddr, |
271 | getBaseAddrFromBox(loc, boxTyPair, a, rewriter)); |
272 | } else { |
273 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractValueOp>(boxaddr, a, 0); |
274 | } |
275 | return mlir::success(); |
276 | } |
277 | }; |
278 | |
279 | /// Convert `!fir.boxchar_len` to `!llvm.extractvalue` for the 2nd part of the |
280 | /// boxchar. |
281 | struct BoxCharLenOpConversion : public fir::FIROpConversion<fir::BoxCharLenOp> { |
282 | using FIROpConversion::FIROpConversion; |
283 | |
284 | mlir::LogicalResult |
285 | matchAndRewrite(fir::BoxCharLenOp boxCharLen, OpAdaptor adaptor, |
286 | mlir::ConversionPatternRewriter &rewriter) const override { |
287 | mlir::Value boxChar = adaptor.getOperands()[0]; |
288 | mlir::Location loc = boxChar.getLoc(); |
289 | mlir::Type returnValTy = boxCharLen.getResult().getType(); |
290 | |
291 | constexpr int boxcharLenIdx = 1; |
292 | auto len = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, boxChar, |
293 | boxcharLenIdx); |
294 | mlir::Value lenAfterCast = integerCast(loc, rewriter, returnValTy, len); |
295 | rewriter.replaceOp(boxCharLen, lenAfterCast); |
296 | |
297 | return mlir::success(); |
298 | } |
299 | }; |
300 | |
301 | /// Lower `fir.box_dims` to a sequence of operations to extract the requested |
302 | /// dimension information from the boxed value. |
303 | /// Result in a triple set of GEPs and loads. |
304 | struct BoxDimsOpConversion : public fir::FIROpConversion<fir::BoxDimsOp> { |
305 | using FIROpConversion::FIROpConversion; |
306 | |
307 | mlir::LogicalResult |
308 | matchAndRewrite(fir::BoxDimsOp boxdims, OpAdaptor adaptor, |
309 | mlir::ConversionPatternRewriter &rewriter) const override { |
310 | llvm::SmallVector<mlir::Type, 3> resultTypes = { |
311 | convertType(boxdims.getResult(0).getType()), |
312 | convertType(boxdims.getResult(1).getType()), |
313 | convertType(boxdims.getResult(2).getType()), |
314 | }; |
315 | TypePair boxTyPair = getBoxTypePair(boxdims.getVal().getType()); |
316 | auto results = getDimsFromBox(boxdims.getLoc(), resultTypes, boxTyPair, |
317 | adaptor.getOperands()[0], |
318 | adaptor.getOperands()[1], rewriter); |
319 | rewriter.replaceOp(boxdims, results); |
320 | return mlir::success(); |
321 | } |
322 | }; |
323 | |
324 | /// Lower `fir.box_elesize` to a sequence of operations ro extract the size of |
325 | /// an element in the boxed value. |
326 | struct BoxEleSizeOpConversion : public fir::FIROpConversion<fir::BoxEleSizeOp> { |
327 | using FIROpConversion::FIROpConversion; |
328 | |
329 | mlir::LogicalResult |
330 | matchAndRewrite(fir::BoxEleSizeOp boxelesz, OpAdaptor adaptor, |
331 | mlir::ConversionPatternRewriter &rewriter) const override { |
332 | mlir::Value box = adaptor.getOperands()[0]; |
333 | auto loc = boxelesz.getLoc(); |
334 | auto ty = convertType(boxelesz.getType()); |
335 | TypePair boxTyPair = getBoxTypePair(boxelesz.getVal().getType()); |
336 | auto elemSize = getElementSizeFromBox(loc, ty, boxTyPair, box, rewriter); |
337 | rewriter.replaceOp(boxelesz, elemSize); |
338 | return mlir::success(); |
339 | } |
340 | }; |
341 | |
342 | /// Lower `fir.box_isalloc` to a sequence of operations to determine if the |
343 | /// boxed value was from an ALLOCATABLE entity. |
344 | struct BoxIsAllocOpConversion : public fir::FIROpConversion<fir::BoxIsAllocOp> { |
345 | using FIROpConversion::FIROpConversion; |
346 | |
347 | mlir::LogicalResult |
348 | matchAndRewrite(fir::BoxIsAllocOp boxisalloc, OpAdaptor adaptor, |
349 | mlir::ConversionPatternRewriter &rewriter) const override { |
350 | mlir::Value box = adaptor.getOperands()[0]; |
351 | auto loc = boxisalloc.getLoc(); |
352 | TypePair boxTyPair = getBoxTypePair(boxisalloc.getVal().getType()); |
353 | mlir::Value check = |
354 | genBoxAttributeCheck(loc, boxTyPair, box, rewriter, kAttrAllocatable); |
355 | rewriter.replaceOp(boxisalloc, check); |
356 | return mlir::success(); |
357 | } |
358 | }; |
359 | |
360 | /// Lower `fir.box_isarray` to a sequence of operations to determine if the |
361 | /// boxed is an array. |
362 | struct BoxIsArrayOpConversion : public fir::FIROpConversion<fir::BoxIsArrayOp> { |
363 | using FIROpConversion::FIROpConversion; |
364 | |
365 | mlir::LogicalResult |
366 | matchAndRewrite(fir::BoxIsArrayOp boxisarray, OpAdaptor adaptor, |
367 | mlir::ConversionPatternRewriter &rewriter) const override { |
368 | mlir::Value a = adaptor.getOperands()[0]; |
369 | auto loc = boxisarray.getLoc(); |
370 | TypePair boxTyPair = getBoxTypePair(boxisarray.getVal().getType()); |
371 | auto rank = getValueFromBox(loc, boxTyPair, a, rewriter.getI32Type(), |
372 | rewriter, kRankPosInBox); |
373 | auto c0 = genConstantOffset(loc, rewriter, 0); |
374 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
375 | boxisarray, mlir::LLVM::ICmpPredicate::ne, rank, c0); |
376 | return mlir::success(); |
377 | } |
378 | }; |
379 | |
380 | /// Lower `fir.box_isptr` to a sequence of operations to determined if the |
381 | /// boxed value was from a POINTER entity. |
382 | struct BoxIsPtrOpConversion : public fir::FIROpConversion<fir::BoxIsPtrOp> { |
383 | using FIROpConversion::FIROpConversion; |
384 | |
385 | mlir::LogicalResult |
386 | matchAndRewrite(fir::BoxIsPtrOp boxisptr, OpAdaptor adaptor, |
387 | mlir::ConversionPatternRewriter &rewriter) const override { |
388 | mlir::Value box = adaptor.getOperands()[0]; |
389 | auto loc = boxisptr.getLoc(); |
390 | TypePair boxTyPair = getBoxTypePair(boxisptr.getVal().getType()); |
391 | mlir::Value check = |
392 | genBoxAttributeCheck(loc, boxTyPair, box, rewriter, kAttrPointer); |
393 | rewriter.replaceOp(boxisptr, check); |
394 | return mlir::success(); |
395 | } |
396 | }; |
397 | |
398 | /// Lower `fir.box_rank` to the sequence of operation to extract the rank from |
399 | /// the box. |
400 | struct BoxRankOpConversion : public fir::FIROpConversion<fir::BoxRankOp> { |
401 | using FIROpConversion::FIROpConversion; |
402 | |
403 | mlir::LogicalResult |
404 | matchAndRewrite(fir::BoxRankOp boxrank, OpAdaptor adaptor, |
405 | mlir::ConversionPatternRewriter &rewriter) const override { |
406 | mlir::Value a = adaptor.getOperands()[0]; |
407 | auto loc = boxrank.getLoc(); |
408 | mlir::Type ty = convertType(boxrank.getType()); |
409 | TypePair boxTyPair = getBoxTypePair(boxrank.getVal().getType()); |
410 | auto result = |
411 | getValueFromBox(loc, boxTyPair, a, ty, rewriter, kRankPosInBox); |
412 | rewriter.replaceOp(boxrank, result); |
413 | return mlir::success(); |
414 | } |
415 | }; |
416 | |
417 | /// Lower `fir.boxproc_host` operation. Extracts the host pointer from the |
418 | /// boxproc. |
419 | /// TODO: Part of supporting Fortran 2003 procedure pointers. |
420 | struct BoxProcHostOpConversion |
421 | : public fir::FIROpConversion<fir::BoxProcHostOp> { |
422 | using FIROpConversion::FIROpConversion; |
423 | |
424 | mlir::LogicalResult |
425 | matchAndRewrite(fir::BoxProcHostOp boxprochost, OpAdaptor adaptor, |
426 | mlir::ConversionPatternRewriter &rewriter) const override { |
427 | TODO(boxprochost.getLoc(), "fir.boxproc_host codegen" ); |
428 | return mlir::failure(); |
429 | } |
430 | }; |
431 | |
432 | /// Lower `fir.box_tdesc` to the sequence of operations to extract the type |
433 | /// descriptor from the box. |
434 | struct BoxTypeDescOpConversion |
435 | : public fir::FIROpConversion<fir::BoxTypeDescOp> { |
436 | using FIROpConversion::FIROpConversion; |
437 | |
438 | mlir::LogicalResult |
439 | matchAndRewrite(fir::BoxTypeDescOp boxtypedesc, OpAdaptor adaptor, |
440 | mlir::ConversionPatternRewriter &rewriter) const override { |
441 | mlir::Value box = adaptor.getOperands()[0]; |
442 | TypePair boxTyPair = getBoxTypePair(boxtypedesc.getBox().getType()); |
443 | auto typeDescAddr = |
444 | loadTypeDescAddress(boxtypedesc.getLoc(), boxTyPair, box, rewriter); |
445 | rewriter.replaceOp(boxtypedesc, typeDescAddr); |
446 | return mlir::success(); |
447 | } |
448 | }; |
449 | |
450 | /// Lower `fir.box_typecode` to a sequence of operations to extract the type |
451 | /// code in the boxed value. |
452 | struct BoxTypeCodeOpConversion |
453 | : public fir::FIROpConversion<fir::BoxTypeCodeOp> { |
454 | using FIROpConversion::FIROpConversion; |
455 | |
456 | mlir::LogicalResult |
457 | matchAndRewrite(fir::BoxTypeCodeOp op, OpAdaptor adaptor, |
458 | mlir::ConversionPatternRewriter &rewriter) const override { |
459 | mlir::Value box = adaptor.getOperands()[0]; |
460 | auto loc = box.getLoc(); |
461 | auto ty = convertType(op.getType()); |
462 | TypePair boxTyPair = getBoxTypePair(op.getBox().getType()); |
463 | auto typeCode = |
464 | getValueFromBox(loc, boxTyPair, box, ty, rewriter, kTypePosInBox); |
465 | rewriter.replaceOp(op, typeCode); |
466 | return mlir::success(); |
467 | } |
468 | }; |
469 | |
470 | /// Lower `fir.string_lit` to LLVM IR dialect operation. |
471 | struct StringLitOpConversion : public fir::FIROpConversion<fir::StringLitOp> { |
472 | using FIROpConversion::FIROpConversion; |
473 | |
474 | mlir::LogicalResult |
475 | matchAndRewrite(fir::StringLitOp constop, OpAdaptor adaptor, |
476 | mlir::ConversionPatternRewriter &rewriter) const override { |
477 | auto ty = convertType(constop.getType()); |
478 | auto attr = constop.getValue(); |
479 | if (attr.isa<mlir::StringAttr>()) { |
480 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>(constop, ty, attr); |
481 | return mlir::success(); |
482 | } |
483 | |
484 | auto charTy = constop.getType().cast<fir::CharacterType>(); |
485 | unsigned bits = lowerTy().characterBitsize(charTy); |
486 | mlir::Type intTy = rewriter.getIntegerType(bits); |
487 | mlir::Location loc = constop.getLoc(); |
488 | mlir::Value cst = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
489 | if (auto arr = attr.dyn_cast<mlir::DenseElementsAttr>()) { |
490 | cst = rewriter.create<mlir::LLVM::ConstantOp>(loc, ty, arr); |
491 | } else if (auto arr = attr.dyn_cast<mlir::ArrayAttr>()) { |
492 | for (auto a : llvm::enumerate(arr.getValue())) { |
493 | // convert each character to a precise bitsize |
494 | auto elemAttr = mlir::IntegerAttr::get( |
495 | intTy, |
496 | a.value().cast<mlir::IntegerAttr>().getValue().zextOrTrunc(bits)); |
497 | auto elemCst = |
498 | rewriter.create<mlir::LLVM::ConstantOp>(loc, intTy, elemAttr); |
499 | cst = rewriter.create<mlir::LLVM::InsertValueOp>(loc, cst, elemCst, |
500 | a.index()); |
501 | } |
502 | } else { |
503 | return mlir::failure(); |
504 | } |
505 | rewriter.replaceOp(constop, cst); |
506 | return mlir::success(); |
507 | } |
508 | }; |
509 | |
510 | /// `fir.call` -> `llvm.call` |
511 | struct CallOpConversion : public fir::FIROpConversion<fir::CallOp> { |
512 | using FIROpConversion::FIROpConversion; |
513 | |
514 | mlir::LogicalResult |
515 | matchAndRewrite(fir::CallOp call, OpAdaptor adaptor, |
516 | mlir::ConversionPatternRewriter &rewriter) const override { |
517 | llvm::SmallVector<mlir::Type> resultTys; |
518 | for (auto r : call.getResults()) |
519 | resultTys.push_back(convertType(r.getType())); |
520 | // Convert arith::FastMathFlagsAttr to LLVM::FastMathFlagsAttr. |
521 | mlir::arith::AttrConvertFastMathToLLVM<fir::CallOp, mlir::LLVM::CallOp> |
522 | attrConvert(call); |
523 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( |
524 | call, resultTys, adaptor.getOperands(), attrConvert.getAttrs()); |
525 | return mlir::success(); |
526 | } |
527 | }; |
528 | } // namespace |
529 | |
530 | static mlir::Type getComplexEleTy(mlir::Type complex) { |
531 | if (auto cc = complex.dyn_cast<mlir::ComplexType>()) |
532 | return cc.getElementType(); |
533 | return complex.cast<fir::ComplexType>().getElementType(); |
534 | } |
535 | |
536 | namespace { |
537 | /// Compare complex values |
538 | /// |
539 | /// Per 10.1, the only comparisons available are .EQ. (oeq) and .NE. (une). |
540 | /// |
541 | /// For completeness, all other comparison are done on the real component only. |
542 | struct CmpcOpConversion : public fir::FIROpConversion<fir::CmpcOp> { |
543 | using FIROpConversion::FIROpConversion; |
544 | |
545 | mlir::LogicalResult |
546 | matchAndRewrite(fir::CmpcOp cmp, OpAdaptor adaptor, |
547 | mlir::ConversionPatternRewriter &rewriter) const override { |
548 | mlir::ValueRange operands = adaptor.getOperands(); |
549 | mlir::Type resTy = convertType(cmp.getType()); |
550 | mlir::Location loc = cmp.getLoc(); |
551 | mlir::LLVM::FastmathFlags fmf = |
552 | mlir::arith::convertArithFastMathFlagsToLLVM(cmp.getFastmath()); |
553 | mlir::LLVM::FCmpPredicate pred = |
554 | static_cast<mlir::LLVM::FCmpPredicate>(cmp.getPredicate()); |
555 | auto rcp = rewriter.create<mlir::LLVM::FCmpOp>( |
556 | loc, resTy, pred, |
557 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[0], 0), |
558 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[1], 0), fmf); |
559 | auto icp = rewriter.create<mlir::LLVM::FCmpOp>( |
560 | loc, resTy, pred, |
561 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[0], 1), |
562 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[1], 1), fmf); |
563 | llvm::SmallVector<mlir::Value, 2> cp = {rcp, icp}; |
564 | switch (cmp.getPredicate()) { |
565 | case mlir::arith::CmpFPredicate::OEQ: // .EQ. |
566 | rewriter.replaceOpWithNewOp<mlir::LLVM::AndOp>(cmp, resTy, cp); |
567 | break; |
568 | case mlir::arith::CmpFPredicate::UNE: // .NE. |
569 | rewriter.replaceOpWithNewOp<mlir::LLVM::OrOp>(cmp, resTy, cp); |
570 | break; |
571 | default: |
572 | rewriter.replaceOp(cmp, rcp.getResult()); |
573 | break; |
574 | } |
575 | return mlir::success(); |
576 | } |
577 | }; |
578 | |
579 | /// Lower complex constants |
580 | struct ConstcOpConversion : public fir::FIROpConversion<fir::ConstcOp> { |
581 | using FIROpConversion::FIROpConversion; |
582 | |
583 | mlir::LogicalResult |
584 | matchAndRewrite(fir::ConstcOp conc, OpAdaptor, |
585 | mlir::ConversionPatternRewriter &rewriter) const override { |
586 | mlir::Location loc = conc.getLoc(); |
587 | mlir::Type ty = convertType(conc.getType()); |
588 | mlir::Type ety = convertType(getComplexEleTy(conc.getType())); |
589 | auto realPart = rewriter.create<mlir::LLVM::ConstantOp>( |
590 | loc, ety, getValue(conc.getReal())); |
591 | auto imPart = rewriter.create<mlir::LLVM::ConstantOp>( |
592 | loc, ety, getValue(conc.getImaginary())); |
593 | auto undef = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
594 | auto setReal = |
595 | rewriter.create<mlir::LLVM::InsertValueOp>(loc, undef, realPart, 0); |
596 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(conc, setReal, |
597 | imPart, 1); |
598 | return mlir::success(); |
599 | } |
600 | |
601 | inline llvm::APFloat getValue(mlir::Attribute attr) const { |
602 | return attr.cast<fir::RealAttr>().getValue(); |
603 | } |
604 | }; |
605 | |
606 | /// convert value of from-type to value of to-type |
607 | struct ConvertOpConversion : public fir::FIROpConversion<fir::ConvertOp> { |
608 | using FIROpConversion::FIROpConversion; |
609 | |
610 | static bool isFloatingPointTy(mlir::Type ty) { |
611 | return ty.isa<mlir::FloatType>(); |
612 | } |
613 | |
614 | mlir::LogicalResult |
615 | matchAndRewrite(fir::ConvertOp convert, OpAdaptor adaptor, |
616 | mlir::ConversionPatternRewriter &rewriter) const override { |
617 | auto fromFirTy = convert.getValue().getType(); |
618 | auto toFirTy = convert.getRes().getType(); |
619 | auto fromTy = convertType(fromFirTy); |
620 | auto toTy = convertType(toFirTy); |
621 | mlir::Value op0 = adaptor.getOperands()[0]; |
622 | |
623 | if (fromFirTy == toFirTy) { |
624 | rewriter.replaceOp(convert, op0); |
625 | return mlir::success(); |
626 | } |
627 | |
628 | auto loc = convert.getLoc(); |
629 | auto i1Type = mlir::IntegerType::get(convert.getContext(), 1); |
630 | |
631 | if (fromFirTy.isa<fir::LogicalType>() || toFirTy.isa<fir::LogicalType>()) { |
632 | // By specification fir::LogicalType value may be any number, |
633 | // where non-zero value represents .true. and zero value represents |
634 | // .false. |
635 | // |
636 | // integer<->logical conversion requires value normalization. |
637 | // Conversion from wide logical to narrow logical must set the result |
638 | // to non-zero iff the input is non-zero - the easiest way to implement |
639 | // it is to compare the input agains zero and set the result to |
640 | // the canonical 0/1. |
641 | // Conversion from narrow logical to wide logical may be implemented |
642 | // as a zero or sign extension of the input, but it may use value |
643 | // normalization as well. |
644 | if (!fromTy.isa<mlir::IntegerType>() || !toTy.isa<mlir::IntegerType>()) |
645 | return mlir::emitError(loc) |
646 | << "unsupported types for logical conversion: " << fromTy |
647 | << " -> " << toTy; |
648 | |
649 | // Do folding for constant inputs. |
650 | if (auto constVal = fir::getIntIfConstant(op0)) { |
651 | mlir::Value normVal = |
652 | genConstantIndex(loc, toTy, rewriter, *constVal ? 1 : 0); |
653 | rewriter.replaceOp(convert, normVal); |
654 | return mlir::success(); |
655 | } |
656 | |
657 | // If the input is i1, then we can just zero extend it, and |
658 | // the result will be normalized. |
659 | if (fromTy == i1Type) { |
660 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, op0); |
661 | return mlir::success(); |
662 | } |
663 | |
664 | // Compare the input with zero. |
665 | mlir::Value zero = genConstantIndex(loc, fromTy, rewriter, 0); |
666 | auto isTrue = rewriter.create<mlir::LLVM::ICmpOp>( |
667 | loc, mlir::LLVM::ICmpPredicate::ne, op0, zero); |
668 | |
669 | // Zero extend the i1 isTrue result to the required type (unless it is i1 |
670 | // itself). |
671 | if (toTy != i1Type) |
672 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, isTrue); |
673 | else |
674 | rewriter.replaceOp(convert, isTrue.getResult()); |
675 | |
676 | return mlir::success(); |
677 | } |
678 | |
679 | if (fromTy == toTy) { |
680 | rewriter.replaceOp(convert, op0); |
681 | return mlir::success(); |
682 | } |
683 | auto convertFpToFp = [&](mlir::Value val, unsigned fromBits, |
684 | unsigned toBits, mlir::Type toTy) -> mlir::Value { |
685 | if (fromBits == toBits) { |
686 | // TODO: Converting between two floating-point representations with the |
687 | // same bitwidth is not allowed for now. |
688 | mlir::emitError(loc, |
689 | "cannot implicitly convert between two floating-point " |
690 | "representations of the same bitwidth" ); |
691 | return {}; |
692 | } |
693 | if (fromBits > toBits) |
694 | return rewriter.create<mlir::LLVM::FPTruncOp>(loc, toTy, val); |
695 | return rewriter.create<mlir::LLVM::FPExtOp>(loc, toTy, val); |
696 | }; |
697 | // Complex to complex conversion. |
698 | if (fir::isa_complex(fromFirTy) && fir::isa_complex(toFirTy)) { |
699 | // Special case: handle the conversion of a complex such that both the |
700 | // real and imaginary parts are converted together. |
701 | auto ty = convertType(getComplexEleTy(convert.getValue().getType())); |
702 | auto rp = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, op0, 0); |
703 | auto ip = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, op0, 1); |
704 | auto nt = convertType(getComplexEleTy(convert.getRes().getType())); |
705 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(ty); |
706 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(nt); |
707 | auto rc = convertFpToFp(rp, fromBits, toBits, nt); |
708 | auto ic = convertFpToFp(ip, fromBits, toBits, nt); |
709 | auto un = rewriter.create<mlir::LLVM::UndefOp>(loc, toTy); |
710 | auto i1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, un, rc, 0); |
711 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(convert, i1, ic, |
712 | 1); |
713 | return mlir::success(); |
714 | } |
715 | |
716 | // Floating point to floating point conversion. |
717 | if (isFloatingPointTy(fromTy)) { |
718 | if (isFloatingPointTy(toTy)) { |
719 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(fromTy); |
720 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(toTy); |
721 | auto v = convertFpToFp(op0, fromBits, toBits, toTy); |
722 | rewriter.replaceOp(convert, v); |
723 | return mlir::success(); |
724 | } |
725 | if (toTy.isa<mlir::IntegerType>()) { |
726 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPToSIOp>(convert, toTy, op0); |
727 | return mlir::success(); |
728 | } |
729 | } else if (fromTy.isa<mlir::IntegerType>()) { |
730 | // Integer to integer conversion. |
731 | if (toTy.isa<mlir::IntegerType>()) { |
732 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(fromTy); |
733 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(toTy); |
734 | assert(fromBits != toBits); |
735 | if (fromBits > toBits) { |
736 | rewriter.replaceOpWithNewOp<mlir::LLVM::TruncOp>(convert, toTy, op0); |
737 | return mlir::success(); |
738 | } |
739 | if (fromFirTy == i1Type) { |
740 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, op0); |
741 | return mlir::success(); |
742 | } |
743 | rewriter.replaceOpWithNewOp<mlir::LLVM::SExtOp>(convert, toTy, op0); |
744 | return mlir::success(); |
745 | } |
746 | // Integer to floating point conversion. |
747 | if (isFloatingPointTy(toTy)) { |
748 | rewriter.replaceOpWithNewOp<mlir::LLVM::SIToFPOp>(convert, toTy, op0); |
749 | return mlir::success(); |
750 | } |
751 | // Integer to pointer conversion. |
752 | if (toTy.isa<mlir::LLVM::LLVMPointerType>()) { |
753 | rewriter.replaceOpWithNewOp<mlir::LLVM::IntToPtrOp>(convert, toTy, op0); |
754 | return mlir::success(); |
755 | } |
756 | } else if (fromTy.isa<mlir::LLVM::LLVMPointerType>()) { |
757 | // Pointer to integer conversion. |
758 | if (toTy.isa<mlir::IntegerType>()) { |
759 | rewriter.replaceOpWithNewOp<mlir::LLVM::PtrToIntOp>(convert, toTy, op0); |
760 | return mlir::success(); |
761 | } |
762 | // Pointer to pointer conversion. |
763 | if (toTy.isa<mlir::LLVM::LLVMPointerType>()) { |
764 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(convert, toTy, op0); |
765 | return mlir::success(); |
766 | } |
767 | } |
768 | return emitError(loc) << "cannot convert " << fromTy << " to " << toTy; |
769 | } |
770 | }; |
771 | |
772 | /// `fir.type_info` operation has no specific CodeGen. The operation is |
773 | /// only used to carry information during FIR to FIR passes. It may be used |
774 | /// in the future to generate the runtime type info data structures instead |
775 | /// of generating them in lowering. |
776 | struct TypeInfoOpConversion : public fir::FIROpConversion<fir::TypeInfoOp> { |
777 | using FIROpConversion::FIROpConversion; |
778 | |
779 | mlir::LogicalResult |
780 | matchAndRewrite(fir::TypeInfoOp op, OpAdaptor, |
781 | mlir::ConversionPatternRewriter &rewriter) const override { |
782 | rewriter.eraseOp(op); |
783 | return mlir::success(); |
784 | } |
785 | }; |
786 | |
787 | /// `fir.dt_entry` operation has no specific CodeGen. The operation is only used |
788 | /// to carry information during FIR to FIR passes. |
789 | struct DTEntryOpConversion : public fir::FIROpConversion<fir::DTEntryOp> { |
790 | using FIROpConversion::FIROpConversion; |
791 | |
792 | mlir::LogicalResult |
793 | matchAndRewrite(fir::DTEntryOp op, OpAdaptor, |
794 | mlir::ConversionPatternRewriter &rewriter) const override { |
795 | rewriter.eraseOp(op); |
796 | return mlir::success(); |
797 | } |
798 | }; |
799 | |
800 | /// Lower `fir.global_len` operation. |
801 | struct GlobalLenOpConversion : public fir::FIROpConversion<fir::GlobalLenOp> { |
802 | using FIROpConversion::FIROpConversion; |
803 | |
804 | mlir::LogicalResult |
805 | matchAndRewrite(fir::GlobalLenOp globalLen, OpAdaptor adaptor, |
806 | mlir::ConversionPatternRewriter &rewriter) const override { |
807 | TODO(globalLen.getLoc(), "fir.global_len codegen" ); |
808 | return mlir::failure(); |
809 | } |
810 | }; |
811 | |
812 | /// Lower fir.len_param_index |
813 | struct LenParamIndexOpConversion |
814 | : public fir::FIROpConversion<fir::LenParamIndexOp> { |
815 | using FIROpConversion::FIROpConversion; |
816 | |
817 | // FIXME: this should be specialized by the runtime target |
818 | mlir::LogicalResult |
819 | matchAndRewrite(fir::LenParamIndexOp lenp, OpAdaptor, |
820 | mlir::ConversionPatternRewriter &rewriter) const override { |
821 | TODO(lenp.getLoc(), "fir.len_param_index codegen" ); |
822 | } |
823 | }; |
824 | |
825 | /// Convert `!fir.emboxchar<!fir.char<KIND, ?>, #n>` into a sequence of |
826 | /// instructions that generate `!llvm.struct<(ptr<ik>, i64)>`. The 1st element |
827 | /// in this struct is a pointer. Its type is determined from `KIND`. The 2nd |
828 | /// element is the length of the character buffer (`#n`). |
829 | struct EmboxCharOpConversion : public fir::FIROpConversion<fir::EmboxCharOp> { |
830 | using FIROpConversion::FIROpConversion; |
831 | |
832 | mlir::LogicalResult |
833 | matchAndRewrite(fir::EmboxCharOp emboxChar, OpAdaptor adaptor, |
834 | mlir::ConversionPatternRewriter &rewriter) const override { |
835 | mlir::ValueRange operands = adaptor.getOperands(); |
836 | |
837 | mlir::Value charBuffer = operands[0]; |
838 | mlir::Value charBufferLen = operands[1]; |
839 | |
840 | mlir::Location loc = emboxChar.getLoc(); |
841 | mlir::Type llvmStructTy = convertType(emboxChar.getType()); |
842 | auto llvmStruct = rewriter.create<mlir::LLVM::UndefOp>(loc, llvmStructTy); |
843 | |
844 | mlir::Type lenTy = |
845 | llvmStructTy.cast<mlir::LLVM::LLVMStructType>().getBody()[1]; |
846 | mlir::Value lenAfterCast = integerCast(loc, rewriter, lenTy, charBufferLen); |
847 | |
848 | mlir::Type addrTy = |
849 | llvmStructTy.cast<mlir::LLVM::LLVMStructType>().getBody()[0]; |
850 | if (addrTy != charBuffer.getType()) |
851 | charBuffer = |
852 | rewriter.create<mlir::LLVM::BitcastOp>(loc, addrTy, charBuffer); |
853 | |
854 | auto insertBufferOp = rewriter.create<mlir::LLVM::InsertValueOp>( |
855 | loc, llvmStruct, charBuffer, 0); |
856 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
857 | emboxChar, insertBufferOp, lenAfterCast, 1); |
858 | |
859 | return mlir::success(); |
860 | } |
861 | }; |
862 | } // namespace |
863 | |
864 | /// Return the LLVMFuncOp corresponding to the standard malloc call. |
865 | static mlir::SymbolRefAttr |
866 | getMalloc(fir::AllocMemOp op, mlir::ConversionPatternRewriter &rewriter) { |
867 | static constexpr char mallocName[] = "malloc" ; |
868 | auto module = op->getParentOfType<mlir::ModuleOp>(); |
869 | if (auto mallocFunc = module.lookupSymbol<mlir::LLVM::LLVMFuncOp>(mallocName)) |
870 | return mlir::SymbolRefAttr::get(mallocFunc); |
871 | if (auto userMalloc = module.lookupSymbol<mlir::func::FuncOp>(mallocName)) |
872 | return mlir::SymbolRefAttr::get(userMalloc); |
873 | mlir::OpBuilder moduleBuilder( |
874 | op->getParentOfType<mlir::ModuleOp>().getBodyRegion()); |
875 | auto indexType = mlir::IntegerType::get(op.getContext(), 64); |
876 | auto mallocDecl = moduleBuilder.create<mlir::LLVM::LLVMFuncOp>( |
877 | op.getLoc(), mallocName, |
878 | mlir::LLVM::LLVMFunctionType::get(getLlvmPtrType(op.getContext()), |
879 | indexType, |
880 | /*isVarArg=*/false)); |
881 | return mlir::SymbolRefAttr::get(mallocDecl); |
882 | } |
883 | |
884 | /// Helper function for generating the LLVM IR that computes the distance |
885 | /// in bytes between adjacent elements pointed to by a pointer |
886 | /// of type \p ptrTy. The result is returned as a value of \p idxTy integer |
887 | /// type. |
888 | static mlir::Value |
889 | computeElementDistance(mlir::Location loc, mlir::Type llvmObjectType, |
890 | mlir::Type idxTy, |
891 | mlir::ConversionPatternRewriter &rewriter) { |
892 | // Note that we cannot use something like |
893 | // mlir::LLVM::getPrimitiveTypeSizeInBits() for the element type here. For |
894 | // example, it returns 10 bytes for mlir::Float80Type for targets where it |
895 | // occupies 16 bytes. Proper solution is probably to use |
896 | // mlir::DataLayout::getTypeABIAlignment(), but DataLayout is not being set |
897 | // yet (see llvm-project#57230). For the time being use the '(intptr_t)((type |
898 | // *)0 + 1)' trick for all types. The generated instructions are optimized |
899 | // into constant by the first pass of InstCombine, so it should not be a |
900 | // performance issue. |
901 | auto llvmPtrTy = ::getLlvmPtrType(llvmObjectType.getContext()); |
902 | auto nullPtr = rewriter.create<mlir::LLVM::ZeroOp>(loc, llvmPtrTy); |
903 | auto gep = rewriter.create<mlir::LLVM::GEPOp>( |
904 | loc, llvmPtrTy, llvmObjectType, nullPtr, |
905 | llvm::ArrayRef<mlir::LLVM::GEPArg>{1}); |
906 | return rewriter.create<mlir::LLVM::PtrToIntOp>(loc, idxTy, gep); |
907 | } |
908 | |
909 | /// Return value of the stride in bytes between adjacent elements |
910 | /// of LLVM type \p llTy. The result is returned as a value of |
911 | /// \p idxTy integer type. |
912 | static mlir::Value |
913 | genTypeStrideInBytes(mlir::Location loc, mlir::Type idxTy, |
914 | mlir::ConversionPatternRewriter &rewriter, |
915 | mlir::Type llTy) { |
916 | // Create a pointer type and use computeElementDistance(). |
917 | return computeElementDistance(loc, llTy, idxTy, rewriter); |
918 | } |
919 | |
920 | namespace { |
921 | /// Lower a `fir.allocmem` instruction into `llvm.call @malloc` |
922 | struct AllocMemOpConversion : public fir::FIROpConversion<fir::AllocMemOp> { |
923 | using FIROpConversion::FIROpConversion; |
924 | |
925 | mlir::LogicalResult |
926 | matchAndRewrite(fir::AllocMemOp heap, OpAdaptor adaptor, |
927 | mlir::ConversionPatternRewriter &rewriter) const override { |
928 | mlir::Type heapTy = heap.getType(); |
929 | mlir::Location loc = heap.getLoc(); |
930 | auto ity = lowerTy().indexType(); |
931 | mlir::Type dataTy = fir::unwrapRefType(heapTy); |
932 | mlir::Type llvmObjectTy = convertObjectType(dataTy); |
933 | if (fir::isRecordWithTypeParameters(fir::unwrapSequenceType(dataTy))) |
934 | TODO(loc, "fir.allocmem codegen of derived type with length parameters" ); |
935 | mlir::Value size = genTypeSizeInBytes(loc, ity, rewriter, llvmObjectTy); |
936 | if (auto scaleSize = genAllocationScaleSize(heap, ity, rewriter)) |
937 | size = rewriter.create<mlir::LLVM::MulOp>(loc, ity, size, scaleSize); |
938 | for (mlir::Value opnd : adaptor.getOperands()) |
939 | size = rewriter.create<mlir::LLVM::MulOp>( |
940 | loc, ity, size, integerCast(loc, rewriter, ity, opnd)); |
941 | heap->setAttr("callee" , getMalloc(heap, rewriter)); |
942 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( |
943 | heap, ::getLlvmPtrType(heap.getContext()), size, heap->getAttrs()); |
944 | return mlir::success(); |
945 | } |
946 | |
947 | /// Compute the allocation size in bytes of the element type of |
948 | /// \p llTy pointer type. The result is returned as a value of \p idxTy |
949 | /// integer type. |
950 | mlir::Value genTypeSizeInBytes(mlir::Location loc, mlir::Type idxTy, |
951 | mlir::ConversionPatternRewriter &rewriter, |
952 | mlir::Type llTy) const { |
953 | return computeElementDistance(loc, llTy, idxTy, rewriter); |
954 | } |
955 | }; |
956 | } // namespace |
957 | |
958 | /// Return the LLVMFuncOp corresponding to the standard free call. |
959 | static mlir::SymbolRefAttr getFree(fir::FreeMemOp op, |
960 | mlir::ConversionPatternRewriter &rewriter) { |
961 | static constexpr char freeName[] = "free" ; |
962 | auto module = op->getParentOfType<mlir::ModuleOp>(); |
963 | // Check if free already defined in the module. |
964 | if (auto freeFunc = module.lookupSymbol<mlir::LLVM::LLVMFuncOp>(freeName)) |
965 | return mlir::SymbolRefAttr::get(freeFunc); |
966 | if (auto freeDefinedByUser = |
967 | module.lookupSymbol<mlir::func::FuncOp>(freeName)) |
968 | return mlir::SymbolRefAttr::get(freeDefinedByUser); |
969 | // Create llvm declaration for free. |
970 | mlir::OpBuilder moduleBuilder(module.getBodyRegion()); |
971 | auto voidType = mlir::LLVM::LLVMVoidType::get(op.getContext()); |
972 | auto freeDecl = moduleBuilder.create<mlir::LLVM::LLVMFuncOp>( |
973 | rewriter.getUnknownLoc(), freeName, |
974 | mlir::LLVM::LLVMFunctionType::get(voidType, |
975 | getLlvmPtrType(op.getContext()), |
976 | /*isVarArg=*/false)); |
977 | return mlir::SymbolRefAttr::get(freeDecl); |
978 | } |
979 | |
980 | static unsigned getDimension(mlir::LLVM::LLVMArrayType ty) { |
981 | unsigned result = 1; |
982 | for (auto eleTy = ty.getElementType().dyn_cast<mlir::LLVM::LLVMArrayType>(); |
983 | eleTy; |
984 | eleTy = eleTy.getElementType().dyn_cast<mlir::LLVM::LLVMArrayType>()) |
985 | ++result; |
986 | return result; |
987 | } |
988 | |
989 | namespace { |
990 | /// Lower a `fir.freemem` instruction into `llvm.call @free` |
991 | struct FreeMemOpConversion : public fir::FIROpConversion<fir::FreeMemOp> { |
992 | using FIROpConversion::FIROpConversion; |
993 | |
994 | mlir::LogicalResult |
995 | matchAndRewrite(fir::FreeMemOp freemem, OpAdaptor adaptor, |
996 | mlir::ConversionPatternRewriter &rewriter) const override { |
997 | mlir::Location loc = freemem.getLoc(); |
998 | freemem->setAttr("callee" , getFree(freemem, rewriter)); |
999 | rewriter.create<mlir::LLVM::CallOp>(loc, mlir::TypeRange{}, |
1000 | mlir::ValueRange{adaptor.getHeapref()}, |
1001 | freemem->getAttrs()); |
1002 | rewriter.eraseOp(freemem); |
1003 | return mlir::success(); |
1004 | } |
1005 | }; |
1006 | } // namespace |
1007 | |
1008 | // Convert subcomponent array indices from column-major to row-major ordering. |
1009 | static llvm::SmallVector<mlir::Value> |
1010 | convertSubcomponentIndices(mlir::Location loc, mlir::Type eleTy, |
1011 | mlir::ValueRange indices, |
1012 | mlir::Type *retTy = nullptr) { |
1013 | llvm::SmallVector<mlir::Value> result; |
1014 | llvm::SmallVector<mlir::Value> arrayIndices; |
1015 | |
1016 | auto appendArrayIndices = [&] { |
1017 | if (arrayIndices.empty()) |
1018 | return; |
1019 | std::reverse(arrayIndices.begin(), arrayIndices.end()); |
1020 | result.append(arrayIndices.begin(), arrayIndices.end()); |
1021 | arrayIndices.clear(); |
1022 | }; |
1023 | |
1024 | for (mlir::Value index : indices) { |
1025 | // Component indices can be field index to select a component, or array |
1026 | // index, to select an element in an array component. |
1027 | if (auto structTy = mlir::dyn_cast<mlir::LLVM::LLVMStructType>(eleTy)) { |
1028 | std::int64_t cstIndex = getConstantIntValue(index); |
1029 | assert(cstIndex < (int64_t)structTy.getBody().size() && |
1030 | "out-of-bounds struct field index" ); |
1031 | eleTy = structTy.getBody()[cstIndex]; |
1032 | appendArrayIndices(); |
1033 | result.push_back(index); |
1034 | } else if (auto arrayTy = |
1035 | mlir::dyn_cast<mlir::LLVM::LLVMArrayType>(eleTy)) { |
1036 | eleTy = arrayTy.getElementType(); |
1037 | arrayIndices.push_back(index); |
1038 | } else |
1039 | fir::emitFatalError(loc, "Unexpected subcomponent type" ); |
1040 | } |
1041 | appendArrayIndices(); |
1042 | if (retTy) |
1043 | *retTy = eleTy; |
1044 | return result; |
1045 | } |
1046 | |
1047 | /// Common base class for embox to descriptor conversion. |
1048 | template <typename OP> |
1049 | struct EmboxCommonConversion : public fir::FIROpConversion<OP> { |
1050 | using fir::FIROpConversion<OP>::FIROpConversion; |
1051 | using TypePair = typename fir::FIROpConversion<OP>::TypePair; |
1052 | |
1053 | static int getCFIAttr(fir::BaseBoxType boxTy) { |
1054 | auto eleTy = boxTy.getEleTy(); |
1055 | if (eleTy.isa<fir::PointerType>()) |
1056 | return CFI_attribute_pointer; |
1057 | if (eleTy.isa<fir::HeapType>()) |
1058 | return CFI_attribute_allocatable; |
1059 | return CFI_attribute_other; |
1060 | } |
1061 | |
1062 | mlir::Value getCharacterByteSize(mlir::Location loc, |
1063 | mlir::ConversionPatternRewriter &rewriter, |
1064 | fir::CharacterType charTy, |
1065 | mlir::ValueRange lenParams) const { |
1066 | auto i64Ty = mlir::IntegerType::get(rewriter.getContext(), 64); |
1067 | mlir::Value size = |
1068 | genTypeStrideInBytes(loc, i64Ty, rewriter, this->convertType(charTy)); |
1069 | if (charTy.hasConstantLen()) |
1070 | return size; // Length accounted for in the genTypeStrideInBytes GEP. |
1071 | // Otherwise, multiply the single character size by the length. |
1072 | assert(!lenParams.empty()); |
1073 | auto len64 = fir::FIROpConversion<OP>::integerCast(loc, rewriter, i64Ty, |
1074 | lenParams.back()); |
1075 | return rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, size, len64); |
1076 | } |
1077 | |
1078 | // Get the element size and CFI type code of the boxed value. |
1079 | std::tuple<mlir::Value, mlir::Value> getSizeAndTypeCode( |
1080 | mlir::Location loc, mlir::ConversionPatternRewriter &rewriter, |
1081 | mlir::Type boxEleTy, mlir::ValueRange lenParams = {}) const { |
1082 | auto i64Ty = mlir::IntegerType::get(rewriter.getContext(), 64); |
1083 | if (auto eleTy = fir::dyn_cast_ptrEleTy(boxEleTy)) |
1084 | boxEleTy = eleTy; |
1085 | if (auto seqTy = boxEleTy.dyn_cast<fir::SequenceType>()) |
1086 | return getSizeAndTypeCode(loc, rewriter, seqTy.getEleTy(), lenParams); |
1087 | if (boxEleTy.isa<mlir::NoneType>()) // unlimited polymorphic or assumed type |
1088 | return {rewriter.create<mlir::LLVM::ConstantOp>(loc, i64Ty, 0), |
1089 | this->genConstantOffset(loc, rewriter, CFI_type_other)}; |
1090 | mlir::Value typeCodeVal = this->genConstantOffset( |
1091 | loc, rewriter, |
1092 | fir::getTypeCode(boxEleTy, this->lowerTy().getKindMap())); |
1093 | if (fir::isa_integer(boxEleTy) || boxEleTy.dyn_cast<fir::LogicalType>() || |
1094 | fir::isa_real(boxEleTy) || fir::isa_complex(boxEleTy)) |
1095 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, |
1096 | this->convertType(boxEleTy)), |
1097 | typeCodeVal}; |
1098 | if (auto charTy = boxEleTy.dyn_cast<fir::CharacterType>()) |
1099 | return {getCharacterByteSize(loc, rewriter, charTy, lenParams), |
1100 | typeCodeVal}; |
1101 | if (fir::isa_ref_type(boxEleTy)) { |
1102 | auto ptrTy = ::getLlvmPtrType(rewriter.getContext()); |
1103 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, ptrTy), typeCodeVal}; |
1104 | } |
1105 | if (boxEleTy.isa<fir::RecordType>()) |
1106 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, |
1107 | this->convertType(boxEleTy)), |
1108 | typeCodeVal}; |
1109 | fir::emitFatalError(loc, "unhandled type in fir.box code generation" ); |
1110 | } |
1111 | |
1112 | /// Basic pattern to write a field in the descriptor |
1113 | mlir::Value insertField(mlir::ConversionPatternRewriter &rewriter, |
1114 | mlir::Location loc, mlir::Value dest, |
1115 | llvm::ArrayRef<std::int64_t> fldIndexes, |
1116 | mlir::Value value, bool bitcast = false) const { |
1117 | auto boxTy = dest.getType(); |
1118 | auto fldTy = this->getBoxEleTy(boxTy, fldIndexes); |
1119 | if (!bitcast) |
1120 | value = this->integerCast(loc, rewriter, fldTy, value); |
1121 | // bitcast are no-ops with LLVM opaque pointers. |
1122 | return rewriter.create<mlir::LLVM::InsertValueOp>(loc, dest, value, |
1123 | fldIndexes); |
1124 | } |
1125 | |
1126 | inline mlir::Value |
1127 | insertBaseAddress(mlir::ConversionPatternRewriter &rewriter, |
1128 | mlir::Location loc, mlir::Value dest, |
1129 | mlir::Value base) const { |
1130 | return insertField(rewriter, loc, dest, {kAddrPosInBox}, base, |
1131 | /*bitCast=*/true); |
1132 | } |
1133 | |
1134 | inline mlir::Value insertLowerBound(mlir::ConversionPatternRewriter &rewriter, |
1135 | mlir::Location loc, mlir::Value dest, |
1136 | unsigned dim, mlir::Value lb) const { |
1137 | return insertField(rewriter, loc, dest, |
1138 | {kDimsPosInBox, dim, kDimLowerBoundPos}, lb); |
1139 | } |
1140 | |
1141 | inline mlir::Value insertExtent(mlir::ConversionPatternRewriter &rewriter, |
1142 | mlir::Location loc, mlir::Value dest, |
1143 | unsigned dim, mlir::Value extent) const { |
1144 | return insertField(rewriter, loc, dest, {kDimsPosInBox, dim, kDimExtentPos}, |
1145 | extent); |
1146 | } |
1147 | |
1148 | inline mlir::Value insertStride(mlir::ConversionPatternRewriter &rewriter, |
1149 | mlir::Location loc, mlir::Value dest, |
1150 | unsigned dim, mlir::Value stride) const { |
1151 | return insertField(rewriter, loc, dest, {kDimsPosInBox, dim, kDimStridePos}, |
1152 | stride); |
1153 | } |
1154 | |
1155 | /// Get the address of the type descriptor global variable that was created by |
1156 | /// lowering for derived type \p recType. |
1157 | mlir::Value getTypeDescriptor(mlir::ModuleOp mod, |
1158 | mlir::ConversionPatternRewriter &rewriter, |
1159 | mlir::Location loc, |
1160 | fir::RecordType recType) const { |
1161 | std::string name = |
1162 | fir::NameUniquer::getTypeDescriptorName(recType.getName()); |
1163 | mlir::Type llvmPtrTy = ::getLlvmPtrType(mod.getContext()); |
1164 | if (auto global = mod.template lookupSymbol<fir::GlobalOp>(name)) { |
1165 | return rewriter.create<mlir::LLVM::AddressOfOp>(loc, llvmPtrTy, |
1166 | global.getSymName()); |
1167 | } |
1168 | if (auto global = mod.template lookupSymbol<mlir::LLVM::GlobalOp>(name)) { |
1169 | // The global may have already been translated to LLVM. |
1170 | return rewriter.create<mlir::LLVM::AddressOfOp>(loc, llvmPtrTy, |
1171 | global.getSymName()); |
1172 | } |
1173 | // Type info derived types do not have type descriptors since they are the |
1174 | // types defining type descriptors. |
1175 | if (!this->options.ignoreMissingTypeDescriptors && |
1176 | !fir::NameUniquer::belongsToModule( |
1177 | name, Fortran::semantics::typeInfoBuiltinModule)) |
1178 | fir::emitFatalError( |
1179 | loc, "runtime derived type info descriptor was not generated" ); |
1180 | return rewriter.create<mlir::LLVM::ZeroOp>(loc, llvmPtrTy); |
1181 | } |
1182 | |
1183 | mlir::Value populateDescriptor(mlir::Location loc, mlir::ModuleOp mod, |
1184 | fir::BaseBoxType boxTy, mlir::Type inputType, |
1185 | mlir::ConversionPatternRewriter &rewriter, |
1186 | unsigned rank, mlir::Value eleSize, |
1187 | mlir::Value cfiTy, |
1188 | mlir::Value typeDesc) const { |
1189 | auto llvmBoxTy = this->lowerTy().convertBoxTypeAsStruct(boxTy, rank); |
1190 | bool isUnlimitedPolymorphic = fir::isUnlimitedPolymorphicType(boxTy); |
1191 | bool useInputType = fir::isPolymorphicType(boxTy) || isUnlimitedPolymorphic; |
1192 | mlir::Value descriptor = |
1193 | rewriter.create<mlir::LLVM::UndefOp>(loc, llvmBoxTy); |
1194 | descriptor = |
1195 | insertField(rewriter, loc, descriptor, {kElemLenPosInBox}, eleSize); |
1196 | descriptor = insertField(rewriter, loc, descriptor, {kVersionPosInBox}, |
1197 | this->genI32Constant(loc, rewriter, CFI_VERSION)); |
1198 | descriptor = insertField(rewriter, loc, descriptor, {kRankPosInBox}, |
1199 | this->genI32Constant(loc, rewriter, rank)); |
1200 | descriptor = insertField(rewriter, loc, descriptor, {kTypePosInBox}, cfiTy); |
1201 | descriptor = |
1202 | insertField(rewriter, loc, descriptor, {kAttributePosInBox}, |
1203 | this->genI32Constant(loc, rewriter, getCFIAttr(boxTy))); |
1204 | const bool hasAddendum = fir::boxHasAddendum(boxTy); |
1205 | descriptor = |
1206 | insertField(rewriter, loc, descriptor, {kF18AddendumPosInBox}, |
1207 | this->genI32Constant(loc, rewriter, hasAddendum ? 1 : 0)); |
1208 | |
1209 | if (hasAddendum) { |
1210 | unsigned typeDescFieldId = getTypeDescFieldId(boxTy); |
1211 | if (!typeDesc) { |
1212 | if (useInputType) { |
1213 | mlir::Type innerType = fir::unwrapInnerType(inputType); |
1214 | if (innerType && innerType.template isa<fir::RecordType>()) { |
1215 | auto recTy = innerType.template dyn_cast<fir::RecordType>(); |
1216 | typeDesc = getTypeDescriptor(mod, rewriter, loc, recTy); |
1217 | } else { |
1218 | // Unlimited polymorphic type descriptor with no record type. Set |
1219 | // type descriptor address to a clean state. |
1220 | typeDesc = rewriter.create<mlir::LLVM::ZeroOp>( |
1221 | loc, ::getLlvmPtrType(mod.getContext())); |
1222 | } |
1223 | } else { |
1224 | typeDesc = getTypeDescriptor(mod, rewriter, loc, |
1225 | fir::unwrapIfDerived(boxTy)); |
1226 | } |
1227 | } |
1228 | if (typeDesc) |
1229 | descriptor = |
1230 | insertField(rewriter, loc, descriptor, {typeDescFieldId}, typeDesc, |
1231 | /*bitCast=*/true); |
1232 | // Always initialize the length parameter field to zero to avoid issues |
1233 | // with uninitialized values in Fortran code trying to compare physical |
1234 | // representation of derived types with pointer/allocatable components. |
1235 | // This has been seen in hashing algorithms using TRANSFER. |
1236 | mlir::Value zero = |
1237 | genConstantIndex(loc, rewriter.getI64Type(), rewriter, 0); |
1238 | descriptor = insertField(rewriter, loc, descriptor, |
1239 | {getLenParamFieldId(boxTy), 0}, zero); |
1240 | } |
1241 | return descriptor; |
1242 | } |
1243 | |
1244 | // Template used for fir::EmboxOp and fir::cg::XEmboxOp |
1245 | template <typename BOX> |
1246 | std::tuple<fir::BaseBoxType, mlir::Value, mlir::Value> |
1247 | consDescriptorPrefix(BOX box, mlir::Type inputType, |
1248 | mlir::ConversionPatternRewriter &rewriter, unsigned rank, |
1249 | [[maybe_unused]] mlir::ValueRange substrParams, |
1250 | mlir::ValueRange lenParams, mlir::Value sourceBox = {}, |
1251 | mlir::Type sourceBoxType = {}) const { |
1252 | auto loc = box.getLoc(); |
1253 | auto boxTy = box.getType().template dyn_cast<fir::BaseBoxType>(); |
1254 | bool useInputType = fir::isPolymorphicType(boxTy) && |
1255 | !fir::isUnlimitedPolymorphicType(inputType); |
1256 | llvm::SmallVector<mlir::Value> typeparams = lenParams; |
1257 | if constexpr (!std::is_same_v<BOX, fir::EmboxOp>) { |
1258 | if (!box.getSubstr().empty() && fir::hasDynamicSize(boxTy.getEleTy())) |
1259 | typeparams.push_back(substrParams[1]); |
1260 | } |
1261 | |
1262 | // Write each of the fields with the appropriate values. |
1263 | // When emboxing an element to a polymorphic descriptor, use the |
1264 | // input type since the destination descriptor type has not the exact |
1265 | // information. |
1266 | auto [eleSize, cfiTy] = getSizeAndTypeCode( |
1267 | loc, rewriter, useInputType ? inputType : boxTy.getEleTy(), typeparams); |
1268 | |
1269 | mlir::Value typeDesc; |
1270 | // When emboxing to a polymorphic box, get the type descriptor, type code |
1271 | // and element size from the source box if any. |
1272 | if (fir::isPolymorphicType(boxTy) && sourceBox) { |
1273 | TypePair sourceBoxTyPair = this->getBoxTypePair(sourceBoxType); |
1274 | typeDesc = |
1275 | this->loadTypeDescAddress(loc, sourceBoxTyPair, sourceBox, rewriter); |
1276 | mlir::Type idxTy = this->lowerTy().indexType(); |
1277 | eleSize = this->getElementSizeFromBox(loc, idxTy, sourceBoxTyPair, |
1278 | sourceBox, rewriter); |
1279 | cfiTy = this->getValueFromBox(loc, sourceBoxTyPair, sourceBox, |
1280 | cfiTy.getType(), rewriter, kTypePosInBox); |
1281 | } |
1282 | auto mod = box->template getParentOfType<mlir::ModuleOp>(); |
1283 | mlir::Value descriptor = populateDescriptor( |
1284 | loc, mod, boxTy, inputType, rewriter, rank, eleSize, cfiTy, typeDesc); |
1285 | |
1286 | return {boxTy, descriptor, eleSize}; |
1287 | } |
1288 | |
1289 | std::tuple<fir::BaseBoxType, mlir::Value, mlir::Value> |
1290 | consDescriptorPrefix(fir::cg::XReboxOp box, mlir::Value loweredBox, |
1291 | mlir::ConversionPatternRewriter &rewriter, unsigned rank, |
1292 | mlir::ValueRange substrParams, |
1293 | mlir::ValueRange lenParams, |
1294 | mlir::Value typeDesc = {}) const { |
1295 | auto loc = box.getLoc(); |
1296 | auto boxTy = box.getType().dyn_cast<fir::BaseBoxType>(); |
1297 | auto inputBoxTy = box.getBox().getType().dyn_cast<fir::BaseBoxType>(); |
1298 | auto inputBoxTyPair = this->getBoxTypePair(inputBoxTy); |
1299 | llvm::SmallVector<mlir::Value> typeparams = lenParams; |
1300 | if (!box.getSubstr().empty() && fir::hasDynamicSize(boxTy.getEleTy())) |
1301 | typeparams.push_back(substrParams[1]); |
1302 | |
1303 | auto [eleSize, cfiTy] = |
1304 | getSizeAndTypeCode(loc, rewriter, boxTy.getEleTy(), typeparams); |
1305 | |
1306 | // Reboxing to a polymorphic entity. eleSize and type code need to |
1307 | // be retrieved from the initial box and propagated to the new box. |
1308 | // If the initial box has an addendum, the type desc must be propagated as |
1309 | // well. |
1310 | if (fir::isPolymorphicType(boxTy)) { |
1311 | mlir::Type idxTy = this->lowerTy().indexType(); |
1312 | eleSize = this->getElementSizeFromBox(loc, idxTy, inputBoxTyPair, |
1313 | loweredBox, rewriter); |
1314 | cfiTy = this->getValueFromBox(loc, inputBoxTyPair, loweredBox, |
1315 | cfiTy.getType(), rewriter, kTypePosInBox); |
1316 | // TODO: For initial box that are unlimited polymorphic entities, this |
1317 | // code must be made conditional because unlimited polymorphic entities |
1318 | // with intrinsic type spec does not have addendum. |
1319 | if (fir::boxHasAddendum(inputBoxTy)) |
1320 | typeDesc = this->loadTypeDescAddress(loc, inputBoxTyPair, loweredBox, |
1321 | rewriter); |
1322 | } |
1323 | |
1324 | auto mod = box->template getParentOfType<mlir::ModuleOp>(); |
1325 | mlir::Value descriptor = |
1326 | populateDescriptor(loc, mod, boxTy, box.getBox().getType(), rewriter, |
1327 | rank, eleSize, cfiTy, typeDesc); |
1328 | |
1329 | return {boxTy, descriptor, eleSize}; |
1330 | } |
1331 | |
1332 | // Compute the base address of a fir.box given the indices from the slice. |
1333 | // The indices from the "outer" dimensions (every dimension after the first |
1334 | // one (included) that is not a compile time constant) must have been |
1335 | // multiplied with the related extents and added together into \p outerOffset. |
1336 | mlir::Value |
1337 | genBoxOffsetGep(mlir::ConversionPatternRewriter &rewriter, mlir::Location loc, |
1338 | mlir::Value base, mlir::Type llvmBaseObjectType, |
1339 | mlir::Value outerOffset, mlir::ValueRange cstInteriorIndices, |
1340 | mlir::ValueRange componentIndices, |
1341 | std::optional<mlir::Value> substringOffset) const { |
1342 | llvm::SmallVector<mlir::LLVM::GEPArg> gepArgs{outerOffset}; |
1343 | mlir::Type resultTy = llvmBaseObjectType; |
1344 | // Fortran is column major, llvm GEP is row major: reverse the indices here. |
1345 | for (mlir::Value interiorIndex : llvm::reverse(cstInteriorIndices)) { |
1346 | auto arrayTy = resultTy.dyn_cast<mlir::LLVM::LLVMArrayType>(); |
1347 | if (!arrayTy) |
1348 | fir::emitFatalError( |
1349 | loc, |
1350 | "corrupted GEP generated being generated in fir.embox/fir.rebox" ); |
1351 | resultTy = arrayTy.getElementType(); |
1352 | gepArgs.push_back(interiorIndex); |
1353 | } |
1354 | llvm::SmallVector<mlir::Value> gepIndices = |
1355 | convertSubcomponentIndices(loc, resultTy, componentIndices, &resultTy); |
1356 | gepArgs.append(gepIndices.begin(), gepIndices.end()); |
1357 | if (substringOffset) { |
1358 | if (auto arrayTy = resultTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) { |
1359 | gepArgs.push_back(*substringOffset); |
1360 | resultTy = arrayTy.getElementType(); |
1361 | } else { |
1362 | // If the CHARACTER length is dynamic, the whole base type should have |
1363 | // degenerated to an llvm.ptr<i[width]>, and there should not be any |
1364 | // cstInteriorIndices/componentIndices. The substring offset can be |
1365 | // added to the outterOffset since it applies on the same LLVM type. |
1366 | if (gepArgs.size() != 1) |
1367 | fir::emitFatalError(loc, |
1368 | "corrupted substring GEP in fir.embox/fir.rebox" ); |
1369 | mlir::Type outterOffsetTy = gepArgs[0].get<mlir::Value>().getType(); |
1370 | mlir::Value cast = |
1371 | this->integerCast(loc, rewriter, outterOffsetTy, *substringOffset); |
1372 | |
1373 | gepArgs[0] = rewriter.create<mlir::LLVM::AddOp>( |
1374 | loc, outterOffsetTy, gepArgs[0].get<mlir::Value>(), cast); |
1375 | } |
1376 | } |
1377 | mlir::Type llvmPtrTy = ::getLlvmPtrType(resultTy.getContext()); |
1378 | return rewriter.create<mlir::LLVM::GEPOp>( |
1379 | loc, llvmPtrTy, llvmBaseObjectType, base, gepArgs); |
1380 | } |
1381 | |
1382 | template <typename BOX> |
1383 | void |
1384 | getSubcomponentIndices(BOX xbox, mlir::Value memref, |
1385 | mlir::ValueRange operands, |
1386 | mlir::SmallVectorImpl<mlir::Value> &indices) const { |
1387 | // For each field in the path add the offset to base via the args list. |
1388 | // In the most general case, some offsets must be computed since |
1389 | // they are not be known until runtime. |
1390 | if (fir::hasDynamicSize(fir::unwrapSequenceType( |
1391 | fir::unwrapPassByRefType(memref.getType())))) |
1392 | TODO(xbox.getLoc(), |
1393 | "fir.embox codegen dynamic size component in derived type" ); |
1394 | indices.append(operands.begin() + xbox.subcomponentOffset(), |
1395 | operands.begin() + xbox.subcomponentOffset() + |
1396 | xbox.getSubcomponent().size()); |
1397 | } |
1398 | |
1399 | static bool isInGlobalOp(mlir::ConversionPatternRewriter &rewriter) { |
1400 | auto *thisBlock = rewriter.getInsertionBlock(); |
1401 | return thisBlock && |
1402 | mlir::isa<mlir::LLVM::GlobalOp>(thisBlock->getParentOp()); |
1403 | } |
1404 | |
1405 | /// If the embox is not in a globalOp body, allocate storage for the box; |
1406 | /// store the value inside and return the generated alloca. Return the input |
1407 | /// value otherwise. |
1408 | mlir::Value |
1409 | placeInMemoryIfNotGlobalInit(mlir::ConversionPatternRewriter &rewriter, |
1410 | mlir::Location loc, mlir::Type boxTy, |
1411 | mlir::Value boxValue) const { |
1412 | if (isInGlobalOp(rewriter)) |
1413 | return boxValue; |
1414 | mlir::Type llvmBoxTy = boxValue.getType(); |
1415 | auto alloca = this->genAllocaAndAddrCastWithType(loc, llvmBoxTy, |
1416 | defaultAlign, rewriter); |
1417 | auto storeOp = rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, alloca); |
1418 | this->attachTBAATag(storeOp, boxTy, boxTy, nullptr); |
1419 | return alloca; |
1420 | } |
1421 | }; |
1422 | |
1423 | /// Compute the extent of a triplet slice (lb:ub:step). |
1424 | static mlir::Value |
1425 | computeTripletExtent(mlir::ConversionPatternRewriter &rewriter, |
1426 | mlir::Location loc, mlir::Value lb, mlir::Value ub, |
1427 | mlir::Value step, mlir::Value zero, mlir::Type type) { |
1428 | mlir::Value extent = rewriter.create<mlir::LLVM::SubOp>(loc, type, ub, lb); |
1429 | extent = rewriter.create<mlir::LLVM::AddOp>(loc, type, extent, step); |
1430 | extent = rewriter.create<mlir::LLVM::SDivOp>(loc, type, extent, step); |
1431 | // If the resulting extent is negative (`ub-lb` and `step` have different |
1432 | // signs), zero must be returned instead. |
1433 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( |
1434 | loc, mlir::LLVM::ICmpPredicate::sgt, extent, zero); |
1435 | return rewriter.create<mlir::LLVM::SelectOp>(loc, cmp, extent, zero); |
1436 | } |
1437 | |
1438 | /// Create a generic box on a memory reference. This conversions lowers the |
1439 | /// abstract box to the appropriate, initialized descriptor. |
1440 | struct EmboxOpConversion : public EmboxCommonConversion<fir::EmboxOp> { |
1441 | using EmboxCommonConversion::EmboxCommonConversion; |
1442 | |
1443 | mlir::LogicalResult |
1444 | matchAndRewrite(fir::EmboxOp embox, OpAdaptor adaptor, |
1445 | mlir::ConversionPatternRewriter &rewriter) const override { |
1446 | mlir::ValueRange operands = adaptor.getOperands(); |
1447 | mlir::Value sourceBox; |
1448 | mlir::Type sourceBoxType; |
1449 | if (embox.getSourceBox()) { |
1450 | sourceBox = operands[embox.getSourceBoxOffset()]; |
1451 | sourceBoxType = embox.getSourceBox().getType(); |
1452 | } |
1453 | assert(!embox.getShape() && "There should be no dims on this embox op" ); |
1454 | auto [boxTy, dest, eleSize] = consDescriptorPrefix( |
1455 | embox, fir::unwrapRefType(embox.getMemref().getType()), rewriter, |
1456 | /*rank=*/0, /*substrParams=*/mlir::ValueRange{}, |
1457 | adaptor.getTypeparams(), sourceBox, sourceBoxType); |
1458 | dest = insertBaseAddress(rewriter, embox.getLoc(), dest, operands[0]); |
1459 | if (fir::isDerivedTypeWithLenParams(boxTy)) { |
1460 | TODO(embox.getLoc(), |
1461 | "fir.embox codegen of derived with length parameters" ); |
1462 | return mlir::failure(); |
1463 | } |
1464 | auto result = |
1465 | placeInMemoryIfNotGlobalInit(rewriter, embox.getLoc(), boxTy, dest); |
1466 | rewriter.replaceOp(embox, result); |
1467 | return mlir::success(); |
1468 | } |
1469 | }; |
1470 | |
1471 | /// Create a generic box on a memory reference. |
1472 | struct XEmboxOpConversion : public EmboxCommonConversion<fir::cg::XEmboxOp> { |
1473 | using EmboxCommonConversion::EmboxCommonConversion; |
1474 | |
1475 | mlir::LogicalResult |
1476 | matchAndRewrite(fir::cg::XEmboxOp xbox, OpAdaptor adaptor, |
1477 | mlir::ConversionPatternRewriter &rewriter) const override { |
1478 | mlir::ValueRange operands = adaptor.getOperands(); |
1479 | mlir::Value sourceBox; |
1480 | mlir::Type sourceBoxType; |
1481 | if (xbox.getSourceBox()) { |
1482 | sourceBox = operands[xbox.getSourceBoxOffset()]; |
1483 | sourceBoxType = xbox.getSourceBox().getType(); |
1484 | } |
1485 | auto [boxTy, dest, resultEleSize] = consDescriptorPrefix( |
1486 | xbox, fir::unwrapRefType(xbox.getMemref().getType()), rewriter, |
1487 | xbox.getOutRank(), adaptor.getSubstr(), adaptor.getLenParams(), |
1488 | sourceBox, sourceBoxType); |
1489 | // Generate the triples in the dims field of the descriptor |
1490 | auto i64Ty = mlir::IntegerType::get(xbox.getContext(), 64); |
1491 | assert(!xbox.getShape().empty() && "must have a shape" ); |
1492 | unsigned shapeOffset = xbox.shapeOffset(); |
1493 | bool hasShift = !xbox.getShift().empty(); |
1494 | unsigned shiftOffset = xbox.shiftOffset(); |
1495 | bool hasSlice = !xbox.getSlice().empty(); |
1496 | unsigned sliceOffset = xbox.sliceOffset(); |
1497 | mlir::Location loc = xbox.getLoc(); |
1498 | mlir::Value zero = genConstantIndex(loc, i64Ty, rewriter, 0); |
1499 | mlir::Value one = genConstantIndex(loc, i64Ty, rewriter, 1); |
1500 | mlir::Value prevPtrOff = one; |
1501 | mlir::Type eleTy = boxTy.getEleTy(); |
1502 | const unsigned rank = xbox.getRank(); |
1503 | llvm::SmallVector<mlir::Value> cstInteriorIndices; |
1504 | unsigned constRows = 0; |
1505 | mlir::Value ptrOffset = zero; |
1506 | mlir::Type memEleTy = fir::dyn_cast_ptrEleTy(xbox.getMemref().getType()); |
1507 | assert(memEleTy.isa<fir::SequenceType>()); |
1508 | auto seqTy = memEleTy.cast<fir::SequenceType>(); |
1509 | mlir::Type seqEleTy = seqTy.getEleTy(); |
1510 | // Adjust the element scaling factor if the element is a dependent type. |
1511 | if (fir::hasDynamicSize(seqEleTy)) { |
1512 | if (auto charTy = seqEleTy.dyn_cast<fir::CharacterType>()) { |
1513 | // The GEP pointer type decays to llvm.ptr<i[width]>. |
1514 | // The scaling factor is the runtime value of the length. |
1515 | assert(!adaptor.getLenParams().empty()); |
1516 | prevPtrOff = FIROpConversion::integerCast( |
1517 | loc, rewriter, i64Ty, adaptor.getLenParams().back()); |
1518 | } else if (seqEleTy.isa<fir::RecordType>()) { |
1519 | // prevPtrOff = ; |
1520 | TODO(loc, "generate call to calculate size of PDT" ); |
1521 | } else { |
1522 | fir::emitFatalError(loc, "unexpected dynamic type" ); |
1523 | } |
1524 | } else { |
1525 | constRows = seqTy.getConstantRows(); |
1526 | } |
1527 | |
1528 | const auto hasSubcomp = !xbox.getSubcomponent().empty(); |
1529 | const bool hasSubstr = !xbox.getSubstr().empty(); |
1530 | // Initial element stride that will be use to compute the step in |
1531 | // each dimension. Initially, this is the size of the input element. |
1532 | // Note that when there are no components/substring, the resultEleSize |
1533 | // that was previously computed matches the input element size. |
1534 | mlir::Value prevDimByteStride = resultEleSize; |
1535 | if (hasSubcomp) { |
1536 | // We have a subcomponent. The step value needs to be the number of |
1537 | // bytes per element (which is a derived type). |
1538 | prevDimByteStride = |
1539 | genTypeStrideInBytes(loc, i64Ty, rewriter, convertType(seqEleTy)); |
1540 | } else if (hasSubstr) { |
1541 | // We have a substring. The step value needs to be the number of bytes |
1542 | // per CHARACTER element. |
1543 | auto charTy = seqEleTy.cast<fir::CharacterType>(); |
1544 | if (fir::hasDynamicSize(charTy)) { |
1545 | prevDimByteStride = |
1546 | getCharacterByteSize(loc, rewriter, charTy, adaptor.getLenParams()); |
1547 | } else { |
1548 | prevDimByteStride = genConstantIndex( |
1549 | loc, i64Ty, rewriter, |
1550 | charTy.getLen() * lowerTy().characterBitsize(charTy) / 8); |
1551 | } |
1552 | } |
1553 | |
1554 | // Process the array subspace arguments (shape, shift, etc.), if any, |
1555 | // translating everything to values in the descriptor wherever the entity |
1556 | // has a dynamic array dimension. |
1557 | for (unsigned di = 0, descIdx = 0; di < rank; ++di) { |
1558 | mlir::Value extent = operands[shapeOffset]; |
1559 | mlir::Value outerExtent = extent; |
1560 | bool skipNext = false; |
1561 | if (hasSlice) { |
1562 | mlir::Value off = operands[sliceOffset]; |
1563 | mlir::Value adj = one; |
1564 | if (hasShift) |
1565 | adj = operands[shiftOffset]; |
1566 | auto ao = rewriter.create<mlir::LLVM::SubOp>(loc, i64Ty, off, adj); |
1567 | if (constRows > 0) { |
1568 | cstInteriorIndices.push_back(ao); |
1569 | } else { |
1570 | auto dimOff = |
1571 | rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, ao, prevPtrOff); |
1572 | ptrOffset = |
1573 | rewriter.create<mlir::LLVM::AddOp>(loc, i64Ty, dimOff, ptrOffset); |
1574 | } |
1575 | if (mlir::isa_and_nonnull<fir::UndefOp>( |
1576 | xbox.getSlice()[3 * di + 1].getDefiningOp())) { |
1577 | // This dimension contains a scalar expression in the array slice op. |
1578 | // The dimension is loop invariant, will be dropped, and will not |
1579 | // appear in the descriptor. |
1580 | skipNext = true; |
1581 | } |
1582 | } |
1583 | if (!skipNext) { |
1584 | // store extent |
1585 | if (hasSlice) |
1586 | extent = computeTripletExtent(rewriter, loc, operands[sliceOffset], |
1587 | operands[sliceOffset + 1], |
1588 | operands[sliceOffset + 2], zero, i64Ty); |
1589 | // Lower bound is normalized to 0 for BIND(C) interoperability. |
1590 | mlir::Value lb = zero; |
1591 | const bool isaPointerOrAllocatable = |
1592 | eleTy.isa<fir::PointerType>() || eleTy.isa<fir::HeapType>(); |
1593 | // Lower bound is defaults to 1 for POINTER, ALLOCATABLE, and |
1594 | // denormalized descriptors. |
1595 | if (isaPointerOrAllocatable || !normalizedLowerBound(xbox)) |
1596 | lb = one; |
1597 | // If there is a shifted origin, and no fir.slice, and this is not |
1598 | // a normalized descriptor then use the value from the shift op as |
1599 | // the lower bound. |
1600 | if (hasShift && !(hasSlice || hasSubcomp || hasSubstr) && |
1601 | (isaPointerOrAllocatable || !normalizedLowerBound(xbox))) { |
1602 | lb = operands[shiftOffset]; |
1603 | auto extentIsEmpty = rewriter.create<mlir::LLVM::ICmpOp>( |
1604 | loc, mlir::LLVM::ICmpPredicate::eq, extent, zero); |
1605 | lb = rewriter.create<mlir::LLVM::SelectOp>(loc, extentIsEmpty, one, |
1606 | lb); |
1607 | } |
1608 | dest = insertLowerBound(rewriter, loc, dest, descIdx, lb); |
1609 | |
1610 | dest = insertExtent(rewriter, loc, dest, descIdx, extent); |
1611 | |
1612 | // store step (scaled by shaped extent) |
1613 | mlir::Value step = prevDimByteStride; |
1614 | if (hasSlice) |
1615 | step = rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, step, |
1616 | operands[sliceOffset + 2]); |
1617 | dest = insertStride(rewriter, loc, dest, descIdx, step); |
1618 | ++descIdx; |
1619 | } |
1620 | |
1621 | // compute the stride and offset for the next natural dimension |
1622 | prevDimByteStride = rewriter.create<mlir::LLVM::MulOp>( |
1623 | loc, i64Ty, prevDimByteStride, outerExtent); |
1624 | if (constRows == 0) |
1625 | prevPtrOff = rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, prevPtrOff, |
1626 | outerExtent); |
1627 | else |
1628 | --constRows; |
1629 | |
1630 | // increment iterators |
1631 | ++shapeOffset; |
1632 | if (hasShift) |
1633 | ++shiftOffset; |
1634 | if (hasSlice) |
1635 | sliceOffset += 3; |
1636 | } |
1637 | mlir::Value base = adaptor.getMemref(); |
1638 | if (hasSlice || hasSubcomp || hasSubstr) { |
1639 | // Shift the base address. |
1640 | llvm::SmallVector<mlir::Value> fieldIndices; |
1641 | std::optional<mlir::Value> substringOffset; |
1642 | if (hasSubcomp) |
1643 | getSubcomponentIndices(xbox, xbox.getMemref(), operands, fieldIndices); |
1644 | if (hasSubstr) |
1645 | substringOffset = operands[xbox.substrOffset()]; |
1646 | mlir::Type llvmBaseType = |
1647 | convertType(fir::unwrapRefType(xbox.getMemref().getType())); |
1648 | base = genBoxOffsetGep(rewriter, loc, base, llvmBaseType, ptrOffset, |
1649 | cstInteriorIndices, fieldIndices, substringOffset); |
1650 | } |
1651 | dest = insertBaseAddress(rewriter, loc, dest, base); |
1652 | if (fir::isDerivedTypeWithLenParams(boxTy)) |
1653 | TODO(loc, "fir.embox codegen of derived with length parameters" ); |
1654 | |
1655 | mlir::Value result = |
1656 | placeInMemoryIfNotGlobalInit(rewriter, loc, boxTy, dest); |
1657 | rewriter.replaceOp(xbox, result); |
1658 | return mlir::success(); |
1659 | } |
1660 | |
1661 | /// Return true if `xbox` has a normalized lower bounds attribute. A box value |
1662 | /// that is neither a POINTER nor an ALLOCATABLE should be normalized to a |
1663 | /// zero origin lower bound for interoperability with BIND(C). |
1664 | inline static bool normalizedLowerBound(fir::cg::XEmboxOp xbox) { |
1665 | return xbox->hasAttr(fir::getNormalizedLowerBoundAttrName()); |
1666 | } |
1667 | }; |
1668 | |
1669 | /// Create a new box given a box reference. |
1670 | struct XReboxOpConversion : public EmboxCommonConversion<fir::cg::XReboxOp> { |
1671 | using EmboxCommonConversion::EmboxCommonConversion; |
1672 | |
1673 | mlir::LogicalResult |
1674 | matchAndRewrite(fir::cg::XReboxOp rebox, OpAdaptor adaptor, |
1675 | mlir::ConversionPatternRewriter &rewriter) const override { |
1676 | mlir::Location loc = rebox.getLoc(); |
1677 | mlir::Type idxTy = lowerTy().indexType(); |
1678 | mlir::Value loweredBox = adaptor.getOperands()[0]; |
1679 | mlir::ValueRange operands = adaptor.getOperands(); |
1680 | |
1681 | // Inside a fir.global, the input box was produced as an llvm.struct<> |
1682 | // because objects cannot be handled in memory inside a fir.global body that |
1683 | // must be constant foldable. However, the type translation are not |
1684 | // contextual, so the fir.box<T> type of the operation that produced the |
1685 | // fir.box was translated to an llvm.ptr<llvm.struct<>> and the MLIR pass |
1686 | // manager inserted a builtin.unrealized_conversion_cast that was inserted |
1687 | // and needs to be removed here. |
1688 | if (isInGlobalOp(rewriter)) |
1689 | if (auto unrealizedCast = |
1690 | loweredBox.getDefiningOp<mlir::UnrealizedConversionCastOp>()) |
1691 | loweredBox = unrealizedCast.getInputs()[0]; |
1692 | |
1693 | TypePair inputBoxTyPair = getBoxTypePair(rebox.getBox().getType()); |
1694 | |
1695 | // Create new descriptor and fill its non-shape related data. |
1696 | llvm::SmallVector<mlir::Value, 2> lenParams; |
1697 | mlir::Type inputEleTy = getInputEleTy(rebox); |
1698 | if (auto charTy = inputEleTy.dyn_cast<fir::CharacterType>()) { |
1699 | if (charTy.hasConstantLen()) { |
1700 | mlir::Value len = |
1701 | genConstantIndex(loc, idxTy, rewriter, charTy.getLen()); |
1702 | lenParams.emplace_back(len); |
1703 | } else { |
1704 | mlir::Value len = getElementSizeFromBox(loc, idxTy, inputBoxTyPair, |
1705 | loweredBox, rewriter); |
1706 | if (charTy.getFKind() != 1) { |
1707 | assert(!isInGlobalOp(rewriter) && |
1708 | "character target in global op must have constant length" ); |
1709 | mlir::Value width = |
1710 | genConstantIndex(loc, idxTy, rewriter, charTy.getFKind()); |
1711 | len = rewriter.create<mlir::LLVM::SDivOp>(loc, idxTy, len, width); |
1712 | } |
1713 | lenParams.emplace_back(len); |
1714 | } |
1715 | } else if (auto recTy = inputEleTy.dyn_cast<fir::RecordType>()) { |
1716 | if (recTy.getNumLenParams() != 0) |
1717 | TODO(loc, "reboxing descriptor of derived type with length parameters" ); |
1718 | } |
1719 | |
1720 | // Rebox on polymorphic entities needs to carry over the dynamic type. |
1721 | mlir::Value typeDescAddr; |
1722 | if (inputBoxTyPair.fir.isa<fir::ClassType>() && |
1723 | rebox.getType().isa<fir::ClassType>()) |
1724 | typeDescAddr = |
1725 | loadTypeDescAddress(loc, inputBoxTyPair, loweredBox, rewriter); |
1726 | |
1727 | auto [boxTy, dest, eleSize] = |
1728 | consDescriptorPrefix(rebox, loweredBox, rewriter, rebox.getOutRank(), |
1729 | adaptor.getSubstr(), lenParams, typeDescAddr); |
1730 | |
1731 | // Read input extents, strides, and base address |
1732 | llvm::SmallVector<mlir::Value> inputExtents; |
1733 | llvm::SmallVector<mlir::Value> inputStrides; |
1734 | const unsigned inputRank = rebox.getRank(); |
1735 | for (unsigned dim = 0; dim < inputRank; ++dim) { |
1736 | llvm::SmallVector<mlir::Value, 3> dimInfo = |
1737 | getDimsFromBox(loc, {idxTy, idxTy, idxTy}, inputBoxTyPair, loweredBox, |
1738 | dim, rewriter); |
1739 | inputExtents.emplace_back(dimInfo[1]); |
1740 | inputStrides.emplace_back(dimInfo[2]); |
1741 | } |
1742 | |
1743 | mlir::Value baseAddr = |
1744 | getBaseAddrFromBox(loc, inputBoxTyPair, loweredBox, rewriter); |
1745 | |
1746 | if (!rebox.getSlice().empty() || !rebox.getSubcomponent().empty()) |
1747 | return sliceBox(rebox, boxTy, dest, baseAddr, inputExtents, inputStrides, |
1748 | operands, rewriter); |
1749 | return reshapeBox(rebox, boxTy, dest, baseAddr, inputExtents, inputStrides, |
1750 | operands, rewriter); |
1751 | } |
1752 | |
1753 | private: |
1754 | /// Write resulting shape and base address in descriptor, and replace rebox |
1755 | /// op. |
1756 | mlir::LogicalResult |
1757 | finalizeRebox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, |
1758 | mlir::Value base, mlir::ValueRange lbounds, |
1759 | mlir::ValueRange extents, mlir::ValueRange strides, |
1760 | mlir::ConversionPatternRewriter &rewriter) const { |
1761 | mlir::Location loc = rebox.getLoc(); |
1762 | mlir::Value zero = |
1763 | genConstantIndex(loc, lowerTy().indexType(), rewriter, 0); |
1764 | mlir::Value one = genConstantIndex(loc, lowerTy().indexType(), rewriter, 1); |
1765 | for (auto iter : llvm::enumerate(llvm::zip(extents, strides))) { |
1766 | mlir::Value extent = std::get<0>(iter.value()); |
1767 | unsigned dim = iter.index(); |
1768 | mlir::Value lb = one; |
1769 | if (!lbounds.empty()) { |
1770 | lb = lbounds[dim]; |
1771 | auto extentIsEmpty = rewriter.create<mlir::LLVM::ICmpOp>( |
1772 | loc, mlir::LLVM::ICmpPredicate::eq, extent, zero); |
1773 | lb = rewriter.create<mlir::LLVM::SelectOp>(loc, extentIsEmpty, one, lb); |
1774 | }; |
1775 | dest = insertLowerBound(rewriter, loc, dest, dim, lb); |
1776 | dest = insertExtent(rewriter, loc, dest, dim, extent); |
1777 | dest = insertStride(rewriter, loc, dest, dim, std::get<1>(iter.value())); |
1778 | } |
1779 | dest = insertBaseAddress(rewriter, loc, dest, base); |
1780 | mlir::Value result = |
1781 | placeInMemoryIfNotGlobalInit(rewriter, rebox.getLoc(), destBoxTy, dest); |
1782 | rewriter.replaceOp(rebox, result); |
1783 | return mlir::success(); |
1784 | } |
1785 | |
1786 | // Apply slice given the base address, extents and strides of the input box. |
1787 | mlir::LogicalResult |
1788 | sliceBox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, |
1789 | mlir::Value base, mlir::ValueRange inputExtents, |
1790 | mlir::ValueRange inputStrides, mlir::ValueRange operands, |
1791 | mlir::ConversionPatternRewriter &rewriter) const { |
1792 | mlir::Location loc = rebox.getLoc(); |
1793 | mlir::Type byteTy = ::getI8Type(rebox.getContext()); |
1794 | mlir::Type idxTy = lowerTy().indexType(); |
1795 | mlir::Value zero = genConstantIndex(loc, idxTy, rewriter, 0); |
1796 | // Apply subcomponent and substring shift on base address. |
1797 | if (!rebox.getSubcomponent().empty() || !rebox.getSubstr().empty()) { |
1798 | // Cast to inputEleTy* so that a GEP can be used. |
1799 | mlir::Type inputEleTy = getInputEleTy(rebox); |
1800 | mlir::Type llvmBaseObjectType = convertType(inputEleTy); |
1801 | llvm::SmallVector<mlir::Value> fieldIndices; |
1802 | std::optional<mlir::Value> substringOffset; |
1803 | if (!rebox.getSubcomponent().empty()) |
1804 | getSubcomponentIndices(rebox, rebox.getBox(), operands, fieldIndices); |
1805 | if (!rebox.getSubstr().empty()) |
1806 | substringOffset = operands[rebox.substrOffset()]; |
1807 | base = genBoxOffsetGep(rewriter, loc, base, llvmBaseObjectType, zero, |
1808 | /*cstInteriorIndices=*/std::nullopt, fieldIndices, |
1809 | substringOffset); |
1810 | } |
1811 | |
1812 | if (rebox.getSlice().empty()) |
1813 | // The array section is of the form array[%component][substring], keep |
1814 | // the input array extents and strides. |
1815 | return finalizeRebox(rebox, destBoxTy, dest, base, |
1816 | /*lbounds*/ std::nullopt, inputExtents, inputStrides, |
1817 | rewriter); |
1818 | |
1819 | // The slice is of the form array(i:j:k)[%component]. Compute new extents |
1820 | // and strides. |
1821 | llvm::SmallVector<mlir::Value> slicedExtents; |
1822 | llvm::SmallVector<mlir::Value> slicedStrides; |
1823 | mlir::Value one = genConstantIndex(loc, idxTy, rewriter, 1); |
1824 | const bool sliceHasOrigins = !rebox.getShift().empty(); |
1825 | unsigned sliceOps = rebox.sliceOffset(); |
1826 | unsigned shiftOps = rebox.shiftOffset(); |
1827 | auto strideOps = inputStrides.begin(); |
1828 | const unsigned inputRank = inputStrides.size(); |
1829 | for (unsigned i = 0; i < inputRank; |
1830 | ++i, ++strideOps, ++shiftOps, sliceOps += 3) { |
1831 | mlir::Value sliceLb = |
1832 | integerCast(loc, rewriter, idxTy, operands[sliceOps]); |
1833 | mlir::Value inputStride = *strideOps; // already idxTy |
1834 | // Apply origin shift: base += (lb-shift)*input_stride |
1835 | mlir::Value sliceOrigin = |
1836 | sliceHasOrigins |
1837 | ? integerCast(loc, rewriter, idxTy, operands[shiftOps]) |
1838 | : one; |
1839 | mlir::Value diff = |
1840 | rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, sliceLb, sliceOrigin); |
1841 | mlir::Value offset = |
1842 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, inputStride); |
1843 | // Strides from the fir.box are in bytes. |
1844 | base = genGEP(loc, byteTy, rewriter, base, offset); |
1845 | // Apply upper bound and step if this is a triplet. Otherwise, the |
1846 | // dimension is dropped and no extents/strides are computed. |
1847 | mlir::Value upper = operands[sliceOps + 1]; |
1848 | const bool isTripletSlice = |
1849 | !mlir::isa_and_nonnull<mlir::LLVM::UndefOp>(upper.getDefiningOp()); |
1850 | if (isTripletSlice) { |
1851 | mlir::Value step = |
1852 | integerCast(loc, rewriter, idxTy, operands[sliceOps + 2]); |
1853 | // extent = ub-lb+step/step |
1854 | mlir::Value sliceUb = integerCast(loc, rewriter, idxTy, upper); |
1855 | mlir::Value extent = computeTripletExtent(rewriter, loc, sliceLb, |
1856 | sliceUb, step, zero, idxTy); |
1857 | slicedExtents.emplace_back(extent); |
1858 | // stride = step*input_stride |
1859 | mlir::Value stride = |
1860 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, step, inputStride); |
1861 | slicedStrides.emplace_back(stride); |
1862 | } |
1863 | } |
1864 | return finalizeRebox(rebox, destBoxTy, dest, base, /*lbounds*/ std::nullopt, |
1865 | slicedExtents, slicedStrides, rewriter); |
1866 | } |
1867 | |
1868 | /// Apply a new shape to the data described by a box given the base address, |
1869 | /// extents and strides of the box. |
1870 | mlir::LogicalResult |
1871 | reshapeBox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, |
1872 | mlir::Value base, mlir::ValueRange inputExtents, |
1873 | mlir::ValueRange inputStrides, mlir::ValueRange operands, |
1874 | mlir::ConversionPatternRewriter &rewriter) const { |
1875 | mlir::ValueRange reboxShifts{operands.begin() + rebox.shiftOffset(), |
1876 | operands.begin() + rebox.shiftOffset() + |
1877 | rebox.getShift().size()}; |
1878 | if (rebox.getShape().empty()) { |
1879 | // Only setting new lower bounds. |
1880 | return finalizeRebox(rebox, destBoxTy, dest, base, reboxShifts, |
1881 | inputExtents, inputStrides, rewriter); |
1882 | } |
1883 | |
1884 | mlir::Location loc = rebox.getLoc(); |
1885 | |
1886 | llvm::SmallVector<mlir::Value> newStrides; |
1887 | llvm::SmallVector<mlir::Value> newExtents; |
1888 | mlir::Type idxTy = lowerTy().indexType(); |
1889 | // First stride from input box is kept. The rest is assumed contiguous |
1890 | // (it is not possible to reshape otherwise). If the input is scalar, |
1891 | // which may be OK if all new extents are ones, the stride does not |
1892 | // matter, use one. |
1893 | mlir::Value stride = inputStrides.empty() |
1894 | ? genConstantIndex(loc, idxTy, rewriter, 1) |
1895 | : inputStrides[0]; |
1896 | for (unsigned i = 0; i < rebox.getShape().size(); ++i) { |
1897 | mlir::Value rawExtent = operands[rebox.shapeOffset() + i]; |
1898 | mlir::Value extent = integerCast(loc, rewriter, idxTy, rawExtent); |
1899 | newExtents.emplace_back(extent); |
1900 | newStrides.emplace_back(stride); |
1901 | // nextStride = extent * stride; |
1902 | stride = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, extent, stride); |
1903 | } |
1904 | return finalizeRebox(rebox, destBoxTy, dest, base, reboxShifts, newExtents, |
1905 | newStrides, rewriter); |
1906 | } |
1907 | |
1908 | /// Return scalar element type of the input box. |
1909 | static mlir::Type getInputEleTy(fir::cg::XReboxOp rebox) { |
1910 | auto ty = fir::dyn_cast_ptrOrBoxEleTy(rebox.getBox().getType()); |
1911 | if (auto seqTy = ty.dyn_cast<fir::SequenceType>()) |
1912 | return seqTy.getEleTy(); |
1913 | return ty; |
1914 | } |
1915 | }; |
1916 | |
1917 | /// Lower `fir.emboxproc` operation. Creates a procedure box. |
1918 | /// TODO: Part of supporting Fortran 2003 procedure pointers. |
1919 | struct EmboxProcOpConversion : public fir::FIROpConversion<fir::EmboxProcOp> { |
1920 | using FIROpConversion::FIROpConversion; |
1921 | |
1922 | mlir::LogicalResult |
1923 | matchAndRewrite(fir::EmboxProcOp emboxproc, OpAdaptor adaptor, |
1924 | mlir::ConversionPatternRewriter &rewriter) const override { |
1925 | TODO(emboxproc.getLoc(), "fir.emboxproc codegen" ); |
1926 | return mlir::failure(); |
1927 | } |
1928 | }; |
1929 | |
1930 | // Code shared between insert_value and extract_value Ops. |
1931 | struct ValueOpCommon { |
1932 | // Translate the arguments pertaining to any multidimensional array to |
1933 | // row-major order for LLVM-IR. |
1934 | static void toRowMajor(llvm::SmallVectorImpl<int64_t> &indices, |
1935 | mlir::Type ty) { |
1936 | assert(ty && "type is null" ); |
1937 | const auto end = indices.size(); |
1938 | for (std::remove_const_t<decltype(end)> i = 0; i < end; ++i) { |
1939 | if (auto seq = ty.dyn_cast<mlir::LLVM::LLVMArrayType>()) { |
1940 | const auto dim = getDimension(seq); |
1941 | if (dim > 1) { |
1942 | auto ub = std::min(i + dim, end); |
1943 | std::reverse(indices.begin() + i, indices.begin() + ub); |
1944 | i += dim - 1; |
1945 | } |
1946 | ty = getArrayElementType(seq); |
1947 | } else if (auto st = ty.dyn_cast<mlir::LLVM::LLVMStructType>()) { |
1948 | ty = st.getBody()[indices[i]]; |
1949 | } else { |
1950 | llvm_unreachable("index into invalid type" ); |
1951 | } |
1952 | } |
1953 | } |
1954 | |
1955 | static llvm::SmallVector<int64_t> |
1956 | collectIndices(mlir::ConversionPatternRewriter &rewriter, |
1957 | mlir::ArrayAttr arrAttr) { |
1958 | llvm::SmallVector<int64_t> indices; |
1959 | for (auto i = arrAttr.begin(), e = arrAttr.end(); i != e; ++i) { |
1960 | if (auto intAttr = i->dyn_cast<mlir::IntegerAttr>()) { |
1961 | indices.push_back(Elt: intAttr.getInt()); |
1962 | } else { |
1963 | auto fieldName = i->cast<mlir::StringAttr>().getValue(); |
1964 | ++i; |
1965 | auto ty = i->cast<mlir::TypeAttr>().getValue(); |
1966 | auto index = ty.cast<fir::RecordType>().getFieldIndex(fieldName); |
1967 | indices.push_back(Elt: index); |
1968 | } |
1969 | } |
1970 | return indices; |
1971 | } |
1972 | |
1973 | private: |
1974 | static mlir::Type getArrayElementType(mlir::LLVM::LLVMArrayType ty) { |
1975 | auto eleTy = ty.getElementType(); |
1976 | while (auto arrTy = eleTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) |
1977 | eleTy = arrTy.getElementType(); |
1978 | return eleTy; |
1979 | } |
1980 | }; |
1981 | |
1982 | namespace { |
1983 | /// Extract a subobject value from an ssa-value of aggregate type |
1984 | struct |
1985 | : public fir::FIROpAndTypeConversion<fir::ExtractValueOp>, |
1986 | public ValueOpCommon { |
1987 | using FIROpAndTypeConversion::FIROpAndTypeConversion; |
1988 | |
1989 | mlir::LogicalResult |
1990 | (fir::ExtractValueOp , mlir::Type ty, OpAdaptor adaptor, |
1991 | mlir::ConversionPatternRewriter &rewriter) const override { |
1992 | mlir::ValueRange operands = adaptor.getOperands(); |
1993 | auto indices = collectIndices(rewriter, extractVal.getCoor()); |
1994 | toRowMajor(indices, operands[0].getType()); |
1995 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractValueOp>( |
1996 | extractVal, operands[0], indices); |
1997 | return mlir::success(); |
1998 | } |
1999 | }; |
2000 | |
2001 | /// InsertValue is the generalized instruction for the composition of new |
2002 | /// aggregate type values. |
2003 | struct InsertValueOpConversion |
2004 | : public fir::FIROpAndTypeConversion<fir::InsertValueOp>, |
2005 | public ValueOpCommon { |
2006 | using FIROpAndTypeConversion::FIROpAndTypeConversion; |
2007 | |
2008 | mlir::LogicalResult |
2009 | doRewrite(fir::InsertValueOp insertVal, mlir::Type ty, OpAdaptor adaptor, |
2010 | mlir::ConversionPatternRewriter &rewriter) const override { |
2011 | mlir::ValueRange operands = adaptor.getOperands(); |
2012 | auto indices = collectIndices(rewriter, insertVal.getCoor()); |
2013 | toRowMajor(indices, operands[0].getType()); |
2014 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
2015 | insertVal, operands[0], operands[1], indices); |
2016 | return mlir::success(); |
2017 | } |
2018 | }; |
2019 | |
2020 | /// InsertOnRange inserts a value into a sequence over a range of offsets. |
2021 | struct InsertOnRangeOpConversion |
2022 | : public fir::FIROpAndTypeConversion<fir::InsertOnRangeOp> { |
2023 | using FIROpAndTypeConversion::FIROpAndTypeConversion; |
2024 | |
2025 | // Increments an array of subscripts in a row major fasion. |
2026 | void incrementSubscripts(llvm::ArrayRef<int64_t> dims, |
2027 | llvm::SmallVectorImpl<int64_t> &subscripts) const { |
2028 | for (size_t i = dims.size(); i > 0; --i) { |
2029 | if (++subscripts[i - 1] < dims[i - 1]) { |
2030 | return; |
2031 | } |
2032 | subscripts[i - 1] = 0; |
2033 | } |
2034 | } |
2035 | |
2036 | mlir::LogicalResult |
2037 | doRewrite(fir::InsertOnRangeOp range, mlir::Type ty, OpAdaptor adaptor, |
2038 | mlir::ConversionPatternRewriter &rewriter) const override { |
2039 | |
2040 | llvm::SmallVector<std::int64_t> dims; |
2041 | auto type = adaptor.getOperands()[0].getType(); |
2042 | |
2043 | // Iteratively extract the array dimensions from the type. |
2044 | while (auto t = type.dyn_cast<mlir::LLVM::LLVMArrayType>()) { |
2045 | dims.push_back(Elt: t.getNumElements()); |
2046 | type = t.getElementType(); |
2047 | } |
2048 | |
2049 | llvm::SmallVector<std::int64_t> lBounds; |
2050 | llvm::SmallVector<std::int64_t> uBounds; |
2051 | |
2052 | // Unzip the upper and lower bound and convert to a row major format. |
2053 | mlir::DenseIntElementsAttr coor = range.getCoor(); |
2054 | auto reversedCoor = llvm::reverse(coor.getValues<int64_t>()); |
2055 | for (auto i = reversedCoor.begin(), e = reversedCoor.end(); i != e; ++i) { |
2056 | uBounds.push_back(Elt: *i++); |
2057 | lBounds.push_back(Elt: *i); |
2058 | } |
2059 | |
2060 | auto &subscripts = lBounds; |
2061 | auto loc = range.getLoc(); |
2062 | mlir::Value lastOp = adaptor.getOperands()[0]; |
2063 | mlir::Value insertVal = adaptor.getOperands()[1]; |
2064 | |
2065 | while (subscripts != uBounds) { |
2066 | lastOp = rewriter.create<mlir::LLVM::InsertValueOp>( |
2067 | loc, lastOp, insertVal, subscripts); |
2068 | |
2069 | incrementSubscripts(dims, subscripts); |
2070 | } |
2071 | |
2072 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
2073 | range, lastOp, insertVal, subscripts); |
2074 | |
2075 | return mlir::success(); |
2076 | } |
2077 | }; |
2078 | } // namespace |
2079 | |
2080 | namespace { |
2081 | /// XArrayCoor is the address arithmetic on a dynamically shaped, sliced, |
2082 | /// shifted etc. array. |
2083 | /// (See the static restriction on coordinate_of.) array_coor determines the |
2084 | /// coordinate (location) of a specific element. |
2085 | struct XArrayCoorOpConversion |
2086 | : public fir::FIROpAndTypeConversion<fir::cg::XArrayCoorOp> { |
2087 | using FIROpAndTypeConversion::FIROpAndTypeConversion; |
2088 | |
2089 | mlir::LogicalResult |
2090 | doRewrite(fir::cg::XArrayCoorOp coor, mlir::Type llvmPtrTy, OpAdaptor adaptor, |
2091 | mlir::ConversionPatternRewriter &rewriter) const override { |
2092 | auto loc = coor.getLoc(); |
2093 | mlir::ValueRange operands = adaptor.getOperands(); |
2094 | unsigned rank = coor.getRank(); |
2095 | assert(coor.getIndices().size() == rank); |
2096 | assert(coor.getShape().empty() || coor.getShape().size() == rank); |
2097 | assert(coor.getShift().empty() || coor.getShift().size() == rank); |
2098 | assert(coor.getSlice().empty() || coor.getSlice().size() == 3 * rank); |
2099 | mlir::Type idxTy = lowerTy().indexType(); |
2100 | unsigned indexOffset = coor.indicesOffset(); |
2101 | unsigned shapeOffset = coor.shapeOffset(); |
2102 | unsigned shiftOffset = coor.shiftOffset(); |
2103 | unsigned sliceOffset = coor.sliceOffset(); |
2104 | auto sliceOps = coor.getSlice().begin(); |
2105 | mlir::Value one = genConstantIndex(loc, idxTy, rewriter, 1); |
2106 | mlir::Value prevExt = one; |
2107 | mlir::Value offset = genConstantIndex(loc, idxTy, rewriter, 0); |
2108 | const bool isShifted = !coor.getShift().empty(); |
2109 | const bool isSliced = !coor.getSlice().empty(); |
2110 | const bool baseIsBoxed = coor.getMemref().getType().isa<fir::BaseBoxType>(); |
2111 | TypePair baseBoxTyPair = |
2112 | baseIsBoxed ? getBoxTypePair(coor.getMemref().getType()) : TypePair{}; |
2113 | mlir::LLVM::IntegerOverflowFlags nsw = |
2114 | mlir::LLVM::IntegerOverflowFlags::nsw; |
2115 | |
2116 | // For each dimension of the array, generate the offset calculation. |
2117 | for (unsigned i = 0; i < rank; ++i, ++indexOffset, ++shapeOffset, |
2118 | ++shiftOffset, sliceOffset += 3, sliceOps += 3) { |
2119 | mlir::Value index = |
2120 | integerCast(loc, rewriter, idxTy, operands[indexOffset]); |
2121 | mlir::Value lb = |
2122 | isShifted ? integerCast(loc, rewriter, idxTy, operands[shiftOffset]) |
2123 | : one; |
2124 | mlir::Value step = one; |
2125 | bool normalSlice = isSliced; |
2126 | // Compute zero based index in dimension i of the element, applying |
2127 | // potential triplets and lower bounds. |
2128 | if (isSliced) { |
2129 | mlir::Value originalUb = *(sliceOps + 1); |
2130 | normalSlice = |
2131 | !mlir::isa_and_nonnull<fir::UndefOp>(originalUb.getDefiningOp()); |
2132 | if (normalSlice) |
2133 | step = integerCast(loc, rewriter, idxTy, operands[sliceOffset + 2]); |
2134 | } |
2135 | auto idx = rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, index, lb, nsw); |
2136 | mlir::Value diff = |
2137 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, idx, step, nsw); |
2138 | if (normalSlice) { |
2139 | mlir::Value sliceLb = |
2140 | integerCast(loc, rewriter, idxTy, operands[sliceOffset]); |
2141 | auto adj = |
2142 | rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, sliceLb, lb, nsw); |
2143 | diff = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, diff, adj, nsw); |
2144 | } |
2145 | // Update the offset given the stride and the zero based index `diff` |
2146 | // that was just computed. |
2147 | if (baseIsBoxed) { |
2148 | // Use stride in bytes from the descriptor. |
2149 | mlir::Value stride = |
2150 | getStrideFromBox(loc, baseBoxTyPair, operands[0], i, rewriter); |
2151 | auto sc = |
2152 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, stride, nsw); |
2153 | offset = |
2154 | rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, offset, nsw); |
2155 | } else { |
2156 | // Use stride computed at last iteration. |
2157 | auto sc = |
2158 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, prevExt, nsw); |
2159 | offset = |
2160 | rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, offset, nsw); |
2161 | // Compute next stride assuming contiguity of the base array |
2162 | // (in element number). |
2163 | auto nextExt = integerCast(loc, rewriter, idxTy, operands[shapeOffset]); |
2164 | prevExt = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, prevExt, |
2165 | nextExt, nsw); |
2166 | } |
2167 | } |
2168 | |
2169 | // Add computed offset to the base address. |
2170 | if (baseIsBoxed) { |
2171 | // Working with byte offsets. The base address is read from the fir.box. |
2172 | // and used in i8* GEP to do the pointer arithmetic. |
2173 | mlir::Type byteTy = ::getI8Type(coor.getContext()); |
2174 | mlir::Value base = |
2175 | getBaseAddrFromBox(loc, baseBoxTyPair, operands[0], rewriter); |
2176 | llvm::SmallVector<mlir::LLVM::GEPArg> args{offset}; |
2177 | auto addr = rewriter.create<mlir::LLVM::GEPOp>(loc, llvmPtrTy, byteTy, |
2178 | base, args); |
2179 | if (coor.getSubcomponent().empty()) { |
2180 | rewriter.replaceOp(coor, addr); |
2181 | return mlir::success(); |
2182 | } |
2183 | // Cast the element address from void* to the derived type so that the |
2184 | // derived type members can be addresses via a GEP using the index of |
2185 | // components. |
2186 | mlir::Type elementType = |
2187 | getLlvmObjectTypeFromBoxType(coor.getMemref().getType()); |
2188 | while (auto arrayTy = elementType.dyn_cast<mlir::LLVM::LLVMArrayType>()) |
2189 | elementType = arrayTy.getElementType(); |
2190 | args.clear(); |
2191 | args.push_back(0); |
2192 | if (!coor.getLenParams().empty()) { |
2193 | // If type parameters are present, then we don't want to use a GEPOp |
2194 | // as below, as the LLVM struct type cannot be statically defined. |
2195 | TODO(loc, "derived type with type parameters" ); |
2196 | } |
2197 | llvm::SmallVector<mlir::Value> indices = convertSubcomponentIndices( |
2198 | loc, elementType, |
2199 | operands.slice(coor.subcomponentOffset(), |
2200 | coor.getSubcomponent().size())); |
2201 | args.append(indices.begin(), indices.end()); |
2202 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(coor, llvmPtrTy, |
2203 | elementType, addr, args); |
2204 | return mlir::success(); |
2205 | } |
2206 | |
2207 | // The array was not boxed, so it must be contiguous. offset is therefore an |
2208 | // element offset and the base type is kept in the GEP unless the element |
2209 | // type size is itself dynamic. |
2210 | mlir::Type objectTy = fir::unwrapRefType(coor.getMemref().getType()); |
2211 | mlir::Type eleType = fir::unwrapSequenceType(objectTy); |
2212 | mlir::Type gepObjectType = convertType(eleType); |
2213 | llvm::SmallVector<mlir::LLVM::GEPArg> args; |
2214 | if (coor.getSubcomponent().empty()) { |
2215 | // No subcomponent. |
2216 | if (!coor.getLenParams().empty()) { |
2217 | // Type parameters. Adjust element size explicitly. |
2218 | auto eleTy = fir::dyn_cast_ptrEleTy(coor.getType()); |
2219 | assert(eleTy && "result must be a reference-like type" ); |
2220 | if (fir::characterWithDynamicLen(eleTy)) { |
2221 | assert(coor.getLenParams().size() == 1); |
2222 | auto length = integerCast(loc, rewriter, idxTy, |
2223 | operands[coor.lenParamsOffset()]); |
2224 | offset = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, offset, |
2225 | length, nsw); |
2226 | } else { |
2227 | TODO(loc, "compute size of derived type with type parameters" ); |
2228 | } |
2229 | } |
2230 | args.push_back(offset); |
2231 | } else { |
2232 | // There are subcomponents. |
2233 | args.push_back(offset); |
2234 | llvm::SmallVector<mlir::Value> indices = convertSubcomponentIndices( |
2235 | loc, gepObjectType, |
2236 | operands.slice(coor.subcomponentOffset(), |
2237 | coor.getSubcomponent().size())); |
2238 | args.append(indices.begin(), indices.end()); |
2239 | } |
2240 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
2241 | coor, llvmPtrTy, gepObjectType, adaptor.getMemref(), args); |
2242 | return mlir::success(); |
2243 | } |
2244 | }; |
2245 | } // namespace |
2246 | |
2247 | /// Convert to (memory) reference to a reference to a subobject. |
2248 | /// The coordinate_of op is a Swiss army knife operation that can be used on |
2249 | /// (memory) references to records, arrays, complex, etc. as well as boxes. |
2250 | /// With unboxed arrays, there is the restriction that the array have a static |
2251 | /// shape in all but the last column. |
2252 | struct CoordinateOpConversion |
2253 | : public fir::FIROpAndTypeConversion<fir::CoordinateOp> { |
2254 | using FIROpAndTypeConversion::FIROpAndTypeConversion; |
2255 | |
2256 | mlir::LogicalResult |
2257 | doRewrite(fir::CoordinateOp coor, mlir::Type ty, OpAdaptor adaptor, |
2258 | mlir::ConversionPatternRewriter &rewriter) const override { |
2259 | mlir::ValueRange operands = adaptor.getOperands(); |
2260 | |
2261 | mlir::Location loc = coor.getLoc(); |
2262 | mlir::Value base = operands[0]; |
2263 | mlir::Type baseObjectTy = coor.getBaseType(); |
2264 | mlir::Type objectTy = fir::dyn_cast_ptrOrBoxEleTy(baseObjectTy); |
2265 | assert(objectTy && "fir.coordinate_of expects a reference type" ); |
2266 | mlir::Type llvmObjectTy = convertType(objectTy); |
2267 | |
2268 | // Complex type - basically, extract the real or imaginary part |
2269 | // FIXME: double check why this is done before the fir.box case below. |
2270 | if (fir::isa_complex(objectTy)) { |
2271 | mlir::Value gep = |
2272 | genGEP(loc, llvmObjectTy, rewriter, base, 0, operands[1]); |
2273 | rewriter.replaceOp(coor, gep); |
2274 | return mlir::success(); |
2275 | } |
2276 | |
2277 | // Boxed type - get the base pointer from the box |
2278 | if (baseObjectTy.dyn_cast<fir::BaseBoxType>()) |
2279 | return doRewriteBox(coor, operands, loc, rewriter); |
2280 | |
2281 | // Reference, pointer or a heap type |
2282 | if (baseObjectTy.isa<fir::ReferenceType, fir::PointerType, fir::HeapType>()) |
2283 | return doRewriteRefOrPtr(coor, llvmObjectTy, operands, loc, rewriter); |
2284 | |
2285 | return rewriter.notifyMatchFailure( |
2286 | coor, "fir.coordinate_of base operand has unsupported type" ); |
2287 | } |
2288 | |
2289 | static unsigned getFieldNumber(fir::RecordType ty, mlir::Value op) { |
2290 | return fir::hasDynamicSize(ty) |
2291 | ? op.getDefiningOp() |
2292 | ->getAttrOfType<mlir::IntegerAttr>("field" ) |
2293 | .getInt() |
2294 | : getConstantIntValue(op); |
2295 | } |
2296 | |
2297 | static bool hasSubDimensions(mlir::Type type) { |
2298 | return type.isa<fir::SequenceType, fir::RecordType, mlir::TupleType>(); |
2299 | } |
2300 | |
2301 | /// Check whether this form of `!fir.coordinate_of` is supported. These |
2302 | /// additional checks are required, because we are not yet able to convert |
2303 | /// all valid forms of `!fir.coordinate_of`. |
2304 | /// TODO: Either implement the unsupported cases or extend the verifier |
2305 | /// in FIROps.cpp instead. |
2306 | static bool supportedCoordinate(mlir::Type type, mlir::ValueRange coors) { |
2307 | const std::size_t numOfCoors = coors.size(); |
2308 | std::size_t i = 0; |
2309 | bool subEle = false; |
2310 | bool ptrEle = false; |
2311 | for (; i < numOfCoors; ++i) { |
2312 | mlir::Value nxtOpnd = coors[i]; |
2313 | if (auto arrTy = type.dyn_cast<fir::SequenceType>()) { |
2314 | subEle = true; |
2315 | i += arrTy.getDimension() - 1; |
2316 | type = arrTy.getEleTy(); |
2317 | } else if (auto recTy = type.dyn_cast<fir::RecordType>()) { |
2318 | subEle = true; |
2319 | type = recTy.getType(getFieldNumber(recTy, nxtOpnd)); |
2320 | } else if (auto tupTy = type.dyn_cast<mlir::TupleType>()) { |
2321 | subEle = true; |
2322 | type = tupTy.getType(getConstantIntValue(nxtOpnd)); |
2323 | } else { |
2324 | ptrEle = true; |
2325 | } |
2326 | } |
2327 | if (ptrEle) |
2328 | return (!subEle) && (numOfCoors == 1); |
2329 | return subEle && (i >= numOfCoors); |
2330 | } |
2331 | |
2332 | /// Walk the abstract memory layout and determine if the path traverses any |
2333 | /// array types with unknown shape. Return true iff all the array types have a |
2334 | /// constant shape along the path. |
2335 | static bool arraysHaveKnownShape(mlir::Type type, mlir::ValueRange coors) { |
2336 | for (std::size_t i = 0, sz = coors.size(); i < sz; ++i) { |
2337 | mlir::Value nxtOpnd = coors[i]; |
2338 | if (auto arrTy = type.dyn_cast<fir::SequenceType>()) { |
2339 | if (fir::sequenceWithNonConstantShape(arrTy)) |
2340 | return false; |
2341 | i += arrTy.getDimension() - 1; |
2342 | type = arrTy.getEleTy(); |
2343 | } else if (auto strTy = type.dyn_cast<fir::RecordType>()) { |
2344 | type = strTy.getType(getFieldNumber(strTy, nxtOpnd)); |
2345 | } else if (auto strTy = type.dyn_cast<mlir::TupleType>()) { |
2346 | type = strTy.getType(getConstantIntValue(nxtOpnd)); |
2347 | } else { |
2348 | return true; |
2349 | } |
2350 | } |
2351 | return true; |
2352 | } |
2353 | |
2354 | private: |
2355 | mlir::LogicalResult |
2356 | doRewriteBox(fir::CoordinateOp coor, mlir::ValueRange operands, |
2357 | mlir::Location loc, |
2358 | mlir::ConversionPatternRewriter &rewriter) const { |
2359 | mlir::Type boxObjTy = coor.getBaseType(); |
2360 | assert(boxObjTy.dyn_cast<fir::BaseBoxType>() && "This is not a `fir.box`" ); |
2361 | TypePair boxTyPair = getBoxTypePair(boxObjTy); |
2362 | |
2363 | mlir::Value boxBaseAddr = operands[0]; |
2364 | |
2365 | // 1. SPECIAL CASE (uses `fir.len_param_index`): |
2366 | // %box = ... : !fir.box<!fir.type<derived{len1:i32}>> |
2367 | // %lenp = fir.len_param_index len1, !fir.type<derived{len1:i32}> |
2368 | // %addr = coordinate_of %box, %lenp |
2369 | if (coor.getNumOperands() == 2) { |
2370 | mlir::Operation *coordinateDef = |
2371 | (*coor.getCoor().begin()).getDefiningOp(); |
2372 | if (mlir::isa_and_nonnull<fir::LenParamIndexOp>(coordinateDef)) |
2373 | TODO(loc, |
2374 | "fir.coordinate_of - fir.len_param_index is not supported yet" ); |
2375 | } |
2376 | |
2377 | // 2. GENERAL CASE: |
2378 | // 2.1. (`fir.array`) |
2379 | // %box = ... : !fix.box<!fir.array<?xU>> |
2380 | // %idx = ... : index |
2381 | // %resultAddr = coordinate_of %box, %idx : !fir.ref<U> |
2382 | // 2.2 (`fir.derived`) |
2383 | // %box = ... : !fix.box<!fir.type<derived_type{field_1:i32}>> |
2384 | // %idx = ... : i32 |
2385 | // %resultAddr = coordinate_of %box, %idx : !fir.ref<i32> |
2386 | // 2.3 (`fir.derived` inside `fir.array`) |
2387 | // %box = ... : !fir.box<!fir.array<10 x !fir.type<derived_1{field_1:f32, |
2388 | // field_2:f32}>>> %idx1 = ... : index %idx2 = ... : i32 %resultAddr = |
2389 | // coordinate_of %box, %idx1, %idx2 : !fir.ref<f32> |
2390 | // 2.4. TODO: Either document or disable any other case that the following |
2391 | // implementation might convert. |
2392 | mlir::Value resultAddr = |
2393 | getBaseAddrFromBox(loc, boxTyPair, boxBaseAddr, rewriter); |
2394 | // Component Type |
2395 | auto cpnTy = fir::dyn_cast_ptrOrBoxEleTy(boxObjTy); |
2396 | mlir::Type llvmPtrTy = ::getLlvmPtrType(coor.getContext()); |
2397 | mlir::Type byteTy = ::getI8Type(coor.getContext()); |
2398 | mlir::LLVM::IntegerOverflowFlags nsw = |
2399 | mlir::LLVM::IntegerOverflowFlags::nsw; |
2400 | |
2401 | for (unsigned i = 1, last = operands.size(); i < last; ++i) { |
2402 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { |
2403 | if (i != 1) |
2404 | TODO(loc, "fir.array nested inside other array and/or derived type" ); |
2405 | // Applies byte strides from the box. Ignore lower bound from box |
2406 | // since fir.coordinate_of indexes are zero based. Lowering takes care |
2407 | // of lower bound aspects. This both accounts for dynamically sized |
2408 | // types and non contiguous arrays. |
2409 | auto idxTy = lowerTy().indexType(); |
2410 | mlir::Value off = genConstantIndex(loc, idxTy, rewriter, 0); |
2411 | for (unsigned index = i, lastIndex = i + arrTy.getDimension(); |
2412 | index < lastIndex; ++index) { |
2413 | mlir::Value stride = getStrideFromBox(loc, boxTyPair, operands[0], |
2414 | index - i, rewriter); |
2415 | auto sc = rewriter.create<mlir::LLVM::MulOp>( |
2416 | loc, idxTy, operands[index], stride, nsw); |
2417 | off = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, off, nsw); |
2418 | } |
2419 | resultAddr = rewriter.create<mlir::LLVM::GEPOp>( |
2420 | loc, llvmPtrTy, byteTy, resultAddr, |
2421 | llvm::ArrayRef<mlir::LLVM::GEPArg>{off}); |
2422 | i += arrTy.getDimension() - 1; |
2423 | cpnTy = arrTy.getEleTy(); |
2424 | } else if (auto recTy = cpnTy.dyn_cast<fir::RecordType>()) { |
2425 | mlir::Value nxtOpnd = operands[i]; |
2426 | cpnTy = recTy.getType(getFieldNumber(recTy, nxtOpnd)); |
2427 | auto llvmRecTy = lowerTy().convertType(recTy); |
2428 | resultAddr = rewriter.create<mlir::LLVM::GEPOp>( |
2429 | loc, llvmPtrTy, llvmRecTy, resultAddr, |
2430 | llvm::ArrayRef<mlir::LLVM::GEPArg>{0, nxtOpnd}); |
2431 | } else { |
2432 | fir::emitFatalError(loc, "unexpected type in coordinate_of" ); |
2433 | } |
2434 | } |
2435 | |
2436 | rewriter.replaceOp(coor, resultAddr); |
2437 | return mlir::success(); |
2438 | } |
2439 | |
2440 | mlir::LogicalResult |
2441 | doRewriteRefOrPtr(fir::CoordinateOp coor, mlir::Type llvmObjectTy, |
2442 | mlir::ValueRange operands, mlir::Location loc, |
2443 | mlir::ConversionPatternRewriter &rewriter) const { |
2444 | mlir::Type baseObjectTy = coor.getBaseType(); |
2445 | |
2446 | // Component Type |
2447 | mlir::Type cpnTy = fir::dyn_cast_ptrOrBoxEleTy(baseObjectTy); |
2448 | bool hasSubdimension = hasSubDimensions(cpnTy); |
2449 | bool columnIsDeferred = !hasSubdimension; |
2450 | |
2451 | if (!supportedCoordinate(cpnTy, operands.drop_front(1))) |
2452 | TODO(loc, "unsupported combination of coordinate operands" ); |
2453 | |
2454 | const bool hasKnownShape = |
2455 | arraysHaveKnownShape(cpnTy, operands.drop_front(1)); |
2456 | |
2457 | // If only the column is `?`, then we can simply place the column value in |
2458 | // the 0-th GEP position. |
2459 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { |
2460 | if (!hasKnownShape) { |
2461 | const unsigned sz = arrTy.getDimension(); |
2462 | if (arraysHaveKnownShape(arrTy.getEleTy(), |
2463 | operands.drop_front(1 + sz))) { |
2464 | fir::SequenceType::ShapeRef shape = arrTy.getShape(); |
2465 | bool allConst = true; |
2466 | for (unsigned i = 0; i < sz - 1; ++i) { |
2467 | if (shape[i] < 0) { |
2468 | allConst = false; |
2469 | break; |
2470 | } |
2471 | } |
2472 | if (allConst) |
2473 | columnIsDeferred = true; |
2474 | } |
2475 | } |
2476 | } |
2477 | |
2478 | if (fir::hasDynamicSize(fir::unwrapSequenceType(cpnTy))) |
2479 | return mlir::emitError( |
2480 | loc, "fir.coordinate_of with a dynamic element size is unsupported" ); |
2481 | |
2482 | if (hasKnownShape || columnIsDeferred) { |
2483 | llvm::SmallVector<mlir::LLVM::GEPArg> offs; |
2484 | if (hasKnownShape && hasSubdimension) { |
2485 | offs.push_back(0); |
2486 | } |
2487 | std::optional<int> dims; |
2488 | llvm::SmallVector<mlir::Value> arrIdx; |
2489 | for (std::size_t i = 1, sz = operands.size(); i < sz; ++i) { |
2490 | mlir::Value nxtOpnd = operands[i]; |
2491 | |
2492 | if (!cpnTy) |
2493 | return mlir::emitError(loc, "invalid coordinate/check failed" ); |
2494 | |
2495 | // check if the i-th coordinate relates to an array |
2496 | if (dims) { |
2497 | arrIdx.push_back(nxtOpnd); |
2498 | int dimsLeft = *dims; |
2499 | if (dimsLeft > 1) { |
2500 | dims = dimsLeft - 1; |
2501 | continue; |
2502 | } |
2503 | cpnTy = cpnTy.cast<fir::SequenceType>().getEleTy(); |
2504 | // append array range in reverse (FIR arrays are column-major) |
2505 | offs.append(arrIdx.rbegin(), arrIdx.rend()); |
2506 | arrIdx.clear(); |
2507 | dims.reset(); |
2508 | continue; |
2509 | } |
2510 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { |
2511 | int d = arrTy.getDimension() - 1; |
2512 | if (d > 0) { |
2513 | dims = d; |
2514 | arrIdx.push_back(nxtOpnd); |
2515 | continue; |
2516 | } |
2517 | cpnTy = cpnTy.cast<fir::SequenceType>().getEleTy(); |
2518 | offs.push_back(nxtOpnd); |
2519 | continue; |
2520 | } |
2521 | |
2522 | // check if the i-th coordinate relates to a field |
2523 | if (auto recTy = cpnTy.dyn_cast<fir::RecordType>()) |
2524 | cpnTy = recTy.getType(getFieldNumber(recTy, nxtOpnd)); |
2525 | else if (auto tupTy = cpnTy.dyn_cast<mlir::TupleType>()) |
2526 | cpnTy = tupTy.getType(getConstantIntValue(nxtOpnd)); |
2527 | else |
2528 | cpnTy = nullptr; |
2529 | |
2530 | offs.push_back(nxtOpnd); |
2531 | } |
2532 | if (dims) |
2533 | offs.append(arrIdx.rbegin(), arrIdx.rend()); |
2534 | mlir::Value base = operands[0]; |
2535 | mlir::Value retval = genGEP(loc, llvmObjectTy, rewriter, base, offs); |
2536 | rewriter.replaceOp(coor, retval); |
2537 | return mlir::success(); |
2538 | } |
2539 | |
2540 | return mlir::emitError( |
2541 | loc, "fir.coordinate_of base operand has unsupported type" ); |
2542 | } |
2543 | }; |
2544 | |
2545 | /// Convert `fir.field_index`. The conversion depends on whether the size of |
2546 | /// the record is static or dynamic. |
2547 | struct FieldIndexOpConversion : public fir::FIROpConversion<fir::FieldIndexOp> { |
2548 | using FIROpConversion::FIROpConversion; |
2549 | |
2550 | // NB: most field references should be resolved by this point |
2551 | mlir::LogicalResult |
2552 | matchAndRewrite(fir::FieldIndexOp field, OpAdaptor adaptor, |
2553 | mlir::ConversionPatternRewriter &rewriter) const override { |
2554 | auto recTy = field.getOnType().cast<fir::RecordType>(); |
2555 | unsigned index = recTy.getFieldIndex(field.getFieldId()); |
2556 | |
2557 | if (!fir::hasDynamicSize(recTy)) { |
2558 | // Derived type has compile-time constant layout. Return index of the |
2559 | // component type in the parent type (to be used in GEP). |
2560 | rewriter.replaceOp(field, mlir::ValueRange{genConstantOffset( |
2561 | field.getLoc(), rewriter, index)}); |
2562 | return mlir::success(); |
2563 | } |
2564 | |
2565 | // Derived type has compile-time constant layout. Call the compiler |
2566 | // generated function to determine the byte offset of the field at runtime. |
2567 | // This returns a non-constant. |
2568 | mlir::FlatSymbolRefAttr symAttr = mlir::SymbolRefAttr::get( |
2569 | field.getContext(), getOffsetMethodName(recTy, field.getFieldId())); |
2570 | mlir::NamedAttribute callAttr = rewriter.getNamedAttr("callee" , symAttr); |
2571 | mlir::NamedAttribute fieldAttr = rewriter.getNamedAttr( |
2572 | "field" , mlir::IntegerAttr::get(lowerTy().indexType(), index)); |
2573 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( |
2574 | field, lowerTy().offsetType(), adaptor.getOperands(), |
2575 | llvm::ArrayRef<mlir::NamedAttribute>{callAttr, fieldAttr}); |
2576 | return mlir::success(); |
2577 | } |
2578 | |
2579 | // Re-Construct the name of the compiler generated method that calculates the |
2580 | // offset |
2581 | inline static std::string getOffsetMethodName(fir::RecordType recTy, |
2582 | llvm::StringRef field) { |
2583 | return recTy.getName().str() + "P." + field.str() + ".offset" ; |
2584 | } |
2585 | }; |
2586 | |
2587 | /// Convert `fir.end` |
2588 | struct FirEndOpConversion : public fir::FIROpConversion<fir::FirEndOp> { |
2589 | using FIROpConversion::FIROpConversion; |
2590 | |
2591 | mlir::LogicalResult |
2592 | matchAndRewrite(fir::FirEndOp firEnd, OpAdaptor, |
2593 | mlir::ConversionPatternRewriter &rewriter) const override { |
2594 | TODO(firEnd.getLoc(), "fir.end codegen" ); |
2595 | return mlir::failure(); |
2596 | } |
2597 | }; |
2598 | |
2599 | /// Lower `fir.type_desc` to a global addr. |
2600 | struct TypeDescOpConversion : public fir::FIROpConversion<fir::TypeDescOp> { |
2601 | using FIROpConversion::FIROpConversion; |
2602 | |
2603 | mlir::LogicalResult |
2604 | matchAndRewrite(fir::TypeDescOp typeDescOp, OpAdaptor adaptor, |
2605 | mlir::ConversionPatternRewriter &rewriter) const override { |
2606 | mlir::Type inTy = typeDescOp.getInType(); |
2607 | assert(inTy.isa<fir::RecordType>() && "expecting fir.type" ); |
2608 | auto recordType = inTy.dyn_cast<fir::RecordType>(); |
2609 | auto module = typeDescOp.getOperation()->getParentOfType<mlir::ModuleOp>(); |
2610 | std::string typeDescName = |
2611 | fir::NameUniquer::getTypeDescriptorName(recordType.getName()); |
2612 | auto llvmPtrTy = ::getLlvmPtrType(typeDescOp.getContext()); |
2613 | if (auto global = module.lookupSymbol<mlir::LLVM::GlobalOp>(typeDescName)) { |
2614 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>( |
2615 | typeDescOp, llvmPtrTy, global.getSymName()); |
2616 | return mlir::success(); |
2617 | } else if (auto global = module.lookupSymbol<fir::GlobalOp>(typeDescName)) { |
2618 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>( |
2619 | typeDescOp, llvmPtrTy, global.getSymName()); |
2620 | return mlir::success(); |
2621 | } |
2622 | return mlir::failure(); |
2623 | } |
2624 | }; |
2625 | |
2626 | /// Lower `fir.has_value` operation to `llvm.return` operation. |
2627 | struct HasValueOpConversion : public fir::FIROpConversion<fir::HasValueOp> { |
2628 | using FIROpConversion::FIROpConversion; |
2629 | |
2630 | mlir::LogicalResult |
2631 | matchAndRewrite(fir::HasValueOp op, OpAdaptor adaptor, |
2632 | mlir::ConversionPatternRewriter &rewriter) const override { |
2633 | rewriter.replaceOpWithNewOp<mlir::LLVM::ReturnOp>(op, |
2634 | adaptor.getOperands()); |
2635 | return mlir::success(); |
2636 | } |
2637 | }; |
2638 | |
2639 | #ifndef NDEBUG |
2640 | // Check if attr's type is compatible with ty. |
2641 | // |
2642 | // This is done by comparing attr's element type, converted to LLVM type, |
2643 | // with ty's element type. |
2644 | // |
2645 | // Only integer and floating point (including complex) attributes are |
2646 | // supported. Also, attr is expected to have a TensorType and ty is expected |
2647 | // to be of LLVMArrayType. If any of the previous conditions is false, then |
2648 | // the specified attr and ty are not supported by this function and are |
2649 | // assumed to be compatible. |
2650 | static inline bool attributeTypeIsCompatible(mlir::MLIRContext *ctx, |
2651 | mlir::Attribute attr, |
2652 | mlir::Type ty) { |
2653 | // Get attr's LLVM element type. |
2654 | if (!attr) |
2655 | return true; |
2656 | auto intOrFpEleAttr = mlir::dyn_cast<mlir::DenseIntOrFPElementsAttr>(attr); |
2657 | if (!intOrFpEleAttr) |
2658 | return true; |
2659 | auto tensorTy = mlir::dyn_cast<mlir::TensorType>(intOrFpEleAttr.getType()); |
2660 | if (!tensorTy) |
2661 | return true; |
2662 | mlir::Type attrEleTy = |
2663 | mlir::LLVMTypeConverter(ctx).convertType(tensorTy.getElementType()); |
2664 | |
2665 | // Get ty's element type. |
2666 | auto arrTy = mlir::dyn_cast<mlir::LLVM::LLVMArrayType>(ty); |
2667 | if (!arrTy) |
2668 | return true; |
2669 | mlir::Type eleTy = arrTy.getElementType(); |
2670 | while ((arrTy = mlir::dyn_cast<mlir::LLVM::LLVMArrayType>(eleTy))) |
2671 | eleTy = arrTy.getElementType(); |
2672 | |
2673 | return attrEleTy == eleTy; |
2674 | } |
2675 | #endif |
2676 | |
2677 | /// Lower `fir.global` operation to `llvm.global` operation. |
2678 | /// `fir.insert_on_range` operations are replaced with constant dense attribute |
2679 | /// if they are applied on the full range. |
2680 | struct GlobalOpConversion : public fir::FIROpConversion<fir::GlobalOp> { |
2681 | using FIROpConversion::FIROpConversion; |
2682 | |
2683 | mlir::LogicalResult |
2684 | matchAndRewrite(fir::GlobalOp global, OpAdaptor adaptor, |
2685 | mlir::ConversionPatternRewriter &rewriter) const override { |
2686 | auto tyAttr = convertType(global.getType()); |
2687 | if (auto boxType = mlir::dyn_cast<fir::BaseBoxType>(global.getType())) |
2688 | tyAttr = this->lowerTy().convertBoxTypeAsStruct(boxType); |
2689 | auto loc = global.getLoc(); |
2690 | mlir::Attribute initAttr = global.getInitVal().value_or(mlir::Attribute()); |
2691 | assert(attributeTypeIsCompatible(global.getContext(), initAttr, tyAttr)); |
2692 | auto linkage = convertLinkage(global.getLinkName()); |
2693 | auto isConst = global.getConstant().has_value(); |
2694 | auto g = rewriter.create<mlir::LLVM::GlobalOp>( |
2695 | loc, tyAttr, isConst, linkage, global.getSymName(), initAttr); |
2696 | |
2697 | auto module = global->getParentOfType<mlir::ModuleOp>(); |
2698 | // Add comdat if necessary |
2699 | if (fir::getTargetTriple(module).supportsCOMDAT() && |
2700 | (linkage == mlir::LLVM::Linkage::Linkonce || |
2701 | linkage == mlir::LLVM::Linkage::LinkonceODR)) { |
2702 | addComdat(g, rewriter, module); |
2703 | } |
2704 | |
2705 | // Apply all non-Fir::GlobalOp attributes to the LLVM::GlobalOp, preserving |
2706 | // them; whilst taking care not to apply attributes that are lowered in |
2707 | // other ways. |
2708 | llvm::SmallDenseSet<llvm::StringRef> elidedAttrsSet( |
2709 | global.getAttributeNames().begin(), global.getAttributeNames().end()); |
2710 | for (auto &attr : global->getAttrs()) |
2711 | if (!elidedAttrsSet.contains(attr.getName().strref())) |
2712 | g->setAttr(attr.getName(), attr.getValue()); |
2713 | |
2714 | auto &gr = g.getInitializerRegion(); |
2715 | rewriter.inlineRegionBefore(global.getRegion(), gr, gr.end()); |
2716 | if (!gr.empty()) { |
2717 | // Replace insert_on_range with a constant dense attribute if the |
2718 | // initialization is on the full range. |
2719 | auto insertOnRangeOps = gr.front().getOps<fir::InsertOnRangeOp>(); |
2720 | for (auto insertOp : insertOnRangeOps) { |
2721 | if (isFullRange(insertOp.getCoor(), insertOp.getType())) { |
2722 | auto seqTyAttr = convertType(insertOp.getType()); |
2723 | auto *op = insertOp.getVal().getDefiningOp(); |
2724 | auto constant = mlir::dyn_cast<mlir::arith::ConstantOp>(op); |
2725 | if (!constant) { |
2726 | auto convertOp = mlir::dyn_cast<fir::ConvertOp>(op); |
2727 | if (!convertOp) |
2728 | continue; |
2729 | constant = mlir::cast<mlir::arith::ConstantOp>( |
2730 | convertOp.getValue().getDefiningOp()); |
2731 | } |
2732 | mlir::Type vecType = mlir::VectorType::get( |
2733 | insertOp.getType().getShape(), constant.getType()); |
2734 | auto denseAttr = mlir::DenseElementsAttr::get( |
2735 | vecType.cast<mlir::ShapedType>(), constant.getValue()); |
2736 | rewriter.setInsertionPointAfter(insertOp); |
2737 | rewriter.replaceOpWithNewOp<mlir::arith::ConstantOp>( |
2738 | insertOp, seqTyAttr, denseAttr); |
2739 | } |
2740 | } |
2741 | } |
2742 | rewriter.eraseOp(global); |
2743 | return mlir::success(); |
2744 | } |
2745 | |
2746 | bool isFullRange(mlir::DenseIntElementsAttr indexes, |
2747 | fir::SequenceType seqTy) const { |
2748 | auto extents = seqTy.getShape(); |
2749 | if (indexes.size() / 2 != static_cast<int64_t>(extents.size())) |
2750 | return false; |
2751 | auto cur_index = indexes.value_begin<int64_t>(); |
2752 | for (unsigned i = 0; i < indexes.size(); i += 2) { |
2753 | if (*(cur_index++) != 0) |
2754 | return false; |
2755 | if (*(cur_index++) != extents[i / 2] - 1) |
2756 | return false; |
2757 | } |
2758 | return true; |
2759 | } |
2760 | |
2761 | // TODO: String comparaison should be avoided. Replace linkName with an |
2762 | // enumeration. |
2763 | mlir::LLVM::Linkage |
2764 | convertLinkage(std::optional<llvm::StringRef> optLinkage) const { |
2765 | if (optLinkage) { |
2766 | auto name = *optLinkage; |
2767 | if (name == "internal" ) |
2768 | return mlir::LLVM::Linkage::Internal; |
2769 | if (name == "linkonce" ) |
2770 | return mlir::LLVM::Linkage::Linkonce; |
2771 | if (name == "linkonce_odr" ) |
2772 | return mlir::LLVM::Linkage::LinkonceODR; |
2773 | if (name == "common" ) |
2774 | return mlir::LLVM::Linkage::Common; |
2775 | if (name == "weak" ) |
2776 | return mlir::LLVM::Linkage::Weak; |
2777 | } |
2778 | return mlir::LLVM::Linkage::External; |
2779 | } |
2780 | |
2781 | private: |
2782 | static void addComdat(mlir::LLVM::GlobalOp &global, |
2783 | mlir::ConversionPatternRewriter &rewriter, |
2784 | mlir::ModuleOp &module) { |
2785 | const char *comdatName = "__llvm_comdat" ; |
2786 | mlir::LLVM::ComdatOp comdatOp = |
2787 | module.lookupSymbol<mlir::LLVM::ComdatOp>(comdatName); |
2788 | if (!comdatOp) { |
2789 | comdatOp = |
2790 | rewriter.create<mlir::LLVM::ComdatOp>(module.getLoc(), comdatName); |
2791 | } |
2792 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
2793 | rewriter.setInsertionPointToEnd(&comdatOp.getBody().back()); |
2794 | auto selectorOp = rewriter.create<mlir::LLVM::ComdatSelectorOp>( |
2795 | comdatOp.getLoc(), global.getSymName(), |
2796 | mlir::LLVM::comdat::Comdat::Any); |
2797 | global.setComdatAttr(mlir::SymbolRefAttr::get( |
2798 | rewriter.getContext(), comdatName, |
2799 | mlir::FlatSymbolRefAttr::get(selectorOp.getSymNameAttr()))); |
2800 | } |
2801 | }; |
2802 | |
2803 | /// `fir.load` --> `llvm.load` |
2804 | struct LoadOpConversion : public fir::FIROpConversion<fir::LoadOp> { |
2805 | using FIROpConversion::FIROpConversion; |
2806 | |
2807 | mlir::LogicalResult |
2808 | matchAndRewrite(fir::LoadOp load, OpAdaptor adaptor, |
2809 | mlir::ConversionPatternRewriter &rewriter) const override { |
2810 | mlir::Type llvmLoadTy = convertObjectType(load.getType()); |
2811 | if (auto boxTy = load.getType().dyn_cast<fir::BaseBoxType>()) { |
2812 | // fir.box is a special case because it is considered as an ssa values in |
2813 | // fir, but it is lowered as a pointer to a descriptor. So |
2814 | // fir.ref<fir.box> and fir.box end up being the same llvm types and |
2815 | // loading a fir.ref<fir.box> is implemented as taking a snapshot of the |
2816 | // descriptor value into a new descriptor temp. |
2817 | auto inputBoxStorage = adaptor.getOperands()[0]; |
2818 | mlir::Location loc = load.getLoc(); |
2819 | fir::SequenceType seqTy = fir::unwrapUntilSeqType(boxTy); |
2820 | // fir.box of assumed rank do not have a storage |
2821 | // size that is know at compile time. The copy needs to be runtime driven |
2822 | // depending on the actual dynamic rank or type. |
2823 | if (seqTy && seqTy.hasUnknownShape()) |
2824 | TODO(loc, "loading or assumed rank fir.box" ); |
2825 | auto boxValue = |
2826 | rewriter.create<mlir::LLVM::LoadOp>(loc, llvmLoadTy, inputBoxStorage); |
2827 | if (std::optional<mlir::ArrayAttr> optionalTag = load.getTbaa()) |
2828 | boxValue.setTBAATags(*optionalTag); |
2829 | else |
2830 | attachTBAATag(boxValue, boxTy, boxTy, nullptr); |
2831 | auto newBoxStorage = |
2832 | genAllocaAndAddrCastWithType(loc, llvmLoadTy, defaultAlign, rewriter); |
2833 | auto storeOp = |
2834 | rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, newBoxStorage); |
2835 | attachTBAATag(storeOp, boxTy, boxTy, nullptr); |
2836 | rewriter.replaceOp(load, newBoxStorage); |
2837 | } else { |
2838 | auto loadOp = rewriter.create<mlir::LLVM::LoadOp>( |
2839 | load.getLoc(), llvmLoadTy, adaptor.getOperands(), load->getAttrs()); |
2840 | if (std::optional<mlir::ArrayAttr> optionalTag = load.getTbaa()) |
2841 | loadOp.setTBAATags(*optionalTag); |
2842 | else |
2843 | attachTBAATag(loadOp, load.getType(), load.getType(), nullptr); |
2844 | rewriter.replaceOp(load, loadOp.getResult()); |
2845 | } |
2846 | return mlir::success(); |
2847 | } |
2848 | }; |
2849 | |
2850 | /// Lower `fir.no_reassoc` to LLVM IR dialect. |
2851 | /// TODO: how do we want to enforce this in LLVM-IR? Can we manipulate the fast |
2852 | /// math flags? |
2853 | struct NoReassocOpConversion : public fir::FIROpConversion<fir::NoReassocOp> { |
2854 | using FIROpConversion::FIROpConversion; |
2855 | |
2856 | mlir::LogicalResult |
2857 | matchAndRewrite(fir::NoReassocOp noreassoc, OpAdaptor adaptor, |
2858 | mlir::ConversionPatternRewriter &rewriter) const override { |
2859 | rewriter.replaceOp(noreassoc, adaptor.getOperands()[0]); |
2860 | return mlir::success(); |
2861 | } |
2862 | }; |
2863 | |
2864 | static void genCondBrOp(mlir::Location loc, mlir::Value cmp, mlir::Block *dest, |
2865 | std::optional<mlir::ValueRange> destOps, |
2866 | mlir::ConversionPatternRewriter &rewriter, |
2867 | mlir::Block *newBlock) { |
2868 | if (destOps) |
2869 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, dest, *destOps, newBlock, |
2870 | mlir::ValueRange()); |
2871 | else |
2872 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, dest, newBlock); |
2873 | } |
2874 | |
2875 | template <typename A, typename B> |
2876 | static void genBrOp(A caseOp, mlir::Block *dest, std::optional<B> destOps, |
2877 | mlir::ConversionPatternRewriter &rewriter) { |
2878 | if (destOps) |
2879 | rewriter.replaceOpWithNewOp<mlir::LLVM::BrOp>(caseOp, *destOps, dest); |
2880 | else |
2881 | rewriter.replaceOpWithNewOp<mlir::LLVM::BrOp>(caseOp, std::nullopt, dest); |
2882 | } |
2883 | |
2884 | static void genCaseLadderStep(mlir::Location loc, mlir::Value cmp, |
2885 | mlir::Block *dest, |
2886 | std::optional<mlir::ValueRange> destOps, |
2887 | mlir::ConversionPatternRewriter &rewriter) { |
2888 | auto *thisBlock = rewriter.getInsertionBlock(); |
2889 | auto *newBlock = createBlock(rewriter, dest); |
2890 | rewriter.setInsertionPointToEnd(thisBlock); |
2891 | genCondBrOp(loc, cmp, dest, destOps, rewriter, newBlock); |
2892 | rewriter.setInsertionPointToEnd(newBlock); |
2893 | } |
2894 | |
2895 | /// Conversion of `fir.select_case` |
2896 | /// |
2897 | /// The `fir.select_case` operation is converted to a if-then-else ladder. |
2898 | /// Depending on the case condition type, one or several comparison and |
2899 | /// conditional branching can be generated. |
2900 | /// |
2901 | /// A point value case such as `case(4)`, a lower bound case such as |
2902 | /// `case(5:)` or an upper bound case such as `case(:3)` are converted to a |
2903 | /// simple comparison between the selector value and the constant value in the |
2904 | /// case. The block associated with the case condition is then executed if |
2905 | /// the comparison succeed otherwise it branch to the next block with the |
2906 | /// comparison for the next case conditon. |
2907 | /// |
2908 | /// A closed interval case condition such as `case(7:10)` is converted with a |
2909 | /// first comparison and conditional branching for the lower bound. If |
2910 | /// successful, it branch to a second block with the comparison for the |
2911 | /// upper bound in the same case condition. |
2912 | /// |
2913 | /// TODO: lowering of CHARACTER type cases is not handled yet. |
2914 | struct SelectCaseOpConversion : public fir::FIROpConversion<fir::SelectCaseOp> { |
2915 | using FIROpConversion::FIROpConversion; |
2916 | |
2917 | mlir::LogicalResult |
2918 | matchAndRewrite(fir::SelectCaseOp caseOp, OpAdaptor adaptor, |
2919 | mlir::ConversionPatternRewriter &rewriter) const override { |
2920 | unsigned conds = caseOp.getNumConditions(); |
2921 | llvm::ArrayRef<mlir::Attribute> cases = caseOp.getCases().getValue(); |
2922 | // Type can be CHARACTER, INTEGER, or LOGICAL (C1145) |
2923 | auto ty = caseOp.getSelector().getType(); |
2924 | if (ty.isa<fir::CharacterType>()) { |
2925 | TODO(caseOp.getLoc(), "fir.select_case codegen with character type" ); |
2926 | return mlir::failure(); |
2927 | } |
2928 | mlir::Value selector = caseOp.getSelector(adaptor.getOperands()); |
2929 | auto loc = caseOp.getLoc(); |
2930 | for (unsigned t = 0; t != conds; ++t) { |
2931 | mlir::Block *dest = caseOp.getSuccessor(t); |
2932 | std::optional<mlir::ValueRange> destOps = |
2933 | caseOp.getSuccessorOperands(adaptor.getOperands(), t); |
2934 | std::optional<mlir::ValueRange> cmpOps = |
2935 | *caseOp.getCompareOperands(adaptor.getOperands(), t); |
2936 | mlir::Value caseArg = *(cmpOps.value().begin()); |
2937 | mlir::Attribute attr = cases[t]; |
2938 | if (attr.isa<fir::PointIntervalAttr>()) { |
2939 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( |
2940 | loc, mlir::LLVM::ICmpPredicate::eq, selector, caseArg); |
2941 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); |
2942 | continue; |
2943 | } |
2944 | if (attr.isa<fir::LowerBoundAttr>()) { |
2945 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( |
2946 | loc, mlir::LLVM::ICmpPredicate::sle, caseArg, selector); |
2947 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); |
2948 | continue; |
2949 | } |
2950 | if (attr.isa<fir::UpperBoundAttr>()) { |
2951 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( |
2952 | loc, mlir::LLVM::ICmpPredicate::sle, selector, caseArg); |
2953 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); |
2954 | continue; |
2955 | } |
2956 | if (attr.isa<fir::ClosedIntervalAttr>()) { |
2957 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( |
2958 | loc, mlir::LLVM::ICmpPredicate::sle, caseArg, selector); |
2959 | auto *thisBlock = rewriter.getInsertionBlock(); |
2960 | auto *newBlock1 = createBlock(rewriter, dest); |
2961 | auto *newBlock2 = createBlock(rewriter, dest); |
2962 | rewriter.setInsertionPointToEnd(thisBlock); |
2963 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, newBlock1, newBlock2); |
2964 | rewriter.setInsertionPointToEnd(newBlock1); |
2965 | mlir::Value caseArg0 = *(cmpOps.value().begin() + 1); |
2966 | auto cmp0 = rewriter.create<mlir::LLVM::ICmpOp>( |
2967 | loc, mlir::LLVM::ICmpPredicate::sle, selector, caseArg0); |
2968 | genCondBrOp(loc, cmp0, dest, destOps, rewriter, newBlock2); |
2969 | rewriter.setInsertionPointToEnd(newBlock2); |
2970 | continue; |
2971 | } |
2972 | assert(attr.isa<mlir::UnitAttr>()); |
2973 | assert((t + 1 == conds) && "unit must be last" ); |
2974 | genBrOp(caseOp, dest, destOps, rewriter); |
2975 | } |
2976 | return mlir::success(); |
2977 | } |
2978 | }; |
2979 | |
2980 | template <typename OP> |
2981 | static void selectMatchAndRewrite(const fir::LLVMTypeConverter &lowering, |
2982 | OP select, typename OP::Adaptor adaptor, |
2983 | mlir::ConversionPatternRewriter &rewriter) { |
2984 | unsigned conds = select.getNumConditions(); |
2985 | auto cases = select.getCases().getValue(); |
2986 | mlir::Value selector = adaptor.getSelector(); |
2987 | auto loc = select.getLoc(); |
2988 | assert(conds > 0 && "select must have cases" ); |
2989 | |
2990 | llvm::SmallVector<mlir::Block *> destinations; |
2991 | llvm::SmallVector<mlir::ValueRange> destinationsOperands; |
2992 | mlir::Block *defaultDestination; |
2993 | mlir::ValueRange defaultOperands; |
2994 | llvm::SmallVector<int32_t> caseValues; |
2995 | |
2996 | for (unsigned t = 0; t != conds; ++t) { |
2997 | mlir::Block *dest = select.getSuccessor(t); |
2998 | auto destOps = select.getSuccessorOperands(adaptor.getOperands(), t); |
2999 | const mlir::Attribute &attr = cases[t]; |
3000 | if (auto intAttr = attr.template dyn_cast<mlir::IntegerAttr>()) { |
3001 | destinations.push_back(dest); |
3002 | destinationsOperands.push_back(destOps ? *destOps : mlir::ValueRange{}); |
3003 | caseValues.push_back(Elt: intAttr.getInt()); |
3004 | continue; |
3005 | } |
3006 | assert(attr.template dyn_cast_or_null<mlir::UnitAttr>()); |
3007 | assert((t + 1 == conds) && "unit must be last" ); |
3008 | defaultDestination = dest; |
3009 | defaultOperands = destOps ? *destOps : mlir::ValueRange{}; |
3010 | } |
3011 | |
3012 | // LLVM::SwitchOp takes a i32 type for the selector. |
3013 | if (select.getSelector().getType() != rewriter.getI32Type()) |
3014 | selector = rewriter.create<mlir::LLVM::TruncOp>(loc, rewriter.getI32Type(), |
3015 | selector); |
3016 | |
3017 | rewriter.replaceOpWithNewOp<mlir::LLVM::SwitchOp>( |
3018 | select, selector, |
3019 | /*defaultDestination=*/defaultDestination, |
3020 | /*defaultOperands=*/defaultOperands, |
3021 | /*caseValues=*/caseValues, |
3022 | /*caseDestinations=*/destinations, |
3023 | /*caseOperands=*/destinationsOperands, |
3024 | /*branchWeights=*/llvm::ArrayRef<std::int32_t>()); |
3025 | } |
3026 | |
3027 | /// conversion of fir::SelectOp to an if-then-else ladder |
3028 | struct SelectOpConversion : public fir::FIROpConversion<fir::SelectOp> { |
3029 | using FIROpConversion::FIROpConversion; |
3030 | |
3031 | mlir::LogicalResult |
3032 | matchAndRewrite(fir::SelectOp op, OpAdaptor adaptor, |
3033 | mlir::ConversionPatternRewriter &rewriter) const override { |
3034 | selectMatchAndRewrite<fir::SelectOp>(lowerTy(), op, adaptor, rewriter); |
3035 | return mlir::success(); |
3036 | } |
3037 | }; |
3038 | |
3039 | /// conversion of fir::SelectRankOp to an if-then-else ladder |
3040 | struct SelectRankOpConversion : public fir::FIROpConversion<fir::SelectRankOp> { |
3041 | using FIROpConversion::FIROpConversion; |
3042 | |
3043 | mlir::LogicalResult |
3044 | matchAndRewrite(fir::SelectRankOp op, OpAdaptor adaptor, |
3045 | mlir::ConversionPatternRewriter &rewriter) const override { |
3046 | selectMatchAndRewrite<fir::SelectRankOp>(lowerTy(), op, adaptor, rewriter); |
3047 | return mlir::success(); |
3048 | } |
3049 | }; |
3050 | |
3051 | /// Lower `fir.select_type` to LLVM IR dialect. |
3052 | struct SelectTypeOpConversion : public fir::FIROpConversion<fir::SelectTypeOp> { |
3053 | using FIROpConversion::FIROpConversion; |
3054 | |
3055 | mlir::LogicalResult |
3056 | matchAndRewrite(fir::SelectTypeOp select, OpAdaptor adaptor, |
3057 | mlir::ConversionPatternRewriter &rewriter) const override { |
3058 | mlir::emitError(select.getLoc(), |
3059 | "fir.select_type should have already been converted" ); |
3060 | return mlir::failure(); |
3061 | } |
3062 | }; |
3063 | |
3064 | /// `fir.store` --> `llvm.store` |
3065 | struct StoreOpConversion : public fir::FIROpConversion<fir::StoreOp> { |
3066 | using FIROpConversion::FIROpConversion; |
3067 | |
3068 | mlir::LogicalResult |
3069 | matchAndRewrite(fir::StoreOp store, OpAdaptor adaptor, |
3070 | mlir::ConversionPatternRewriter &rewriter) const override { |
3071 | mlir::Location loc = store.getLoc(); |
3072 | mlir::Type storeTy = store.getValue().getType(); |
3073 | mlir::LLVM::StoreOp newStoreOp; |
3074 | if (auto boxTy = storeTy.dyn_cast<fir::BaseBoxType>()) { |
3075 | // fir.box value is actually in memory, load it first before storing it. |
3076 | mlir::Type llvmBoxTy = lowerTy().convertBoxTypeAsStruct(boxTy); |
3077 | auto val = rewriter.create<mlir::LLVM::LoadOp>(loc, llvmBoxTy, |
3078 | adaptor.getOperands()[0]); |
3079 | attachTBAATag(val, boxTy, boxTy, nullptr); |
3080 | newStoreOp = rewriter.create<mlir::LLVM::StoreOp>( |
3081 | loc, val, adaptor.getOperands()[1]); |
3082 | } else { |
3083 | newStoreOp = rewriter.create<mlir::LLVM::StoreOp>( |
3084 | loc, adaptor.getOperands()[0], adaptor.getOperands()[1]); |
3085 | } |
3086 | if (std::optional<mlir::ArrayAttr> optionalTag = store.getTbaa()) |
3087 | newStoreOp.setTBAATags(*optionalTag); |
3088 | else |
3089 | attachTBAATag(newStoreOp, storeTy, storeTy, nullptr); |
3090 | rewriter.eraseOp(store); |
3091 | return mlir::success(); |
3092 | } |
3093 | }; |
3094 | |
3095 | namespace { |
3096 | |
3097 | /// Convert `fir.unboxchar` into two `llvm.extractvalue` instructions. One for |
3098 | /// the character buffer and one for the buffer length. |
3099 | struct UnboxCharOpConversion : public fir::FIROpConversion<fir::UnboxCharOp> { |
3100 | using FIROpConversion::FIROpConversion; |
3101 | |
3102 | mlir::LogicalResult |
3103 | matchAndRewrite(fir::UnboxCharOp unboxchar, OpAdaptor adaptor, |
3104 | mlir::ConversionPatternRewriter &rewriter) const override { |
3105 | mlir::Type lenTy = convertType(unboxchar.getType(1)); |
3106 | mlir::Value tuple = adaptor.getOperands()[0]; |
3107 | |
3108 | mlir::Location loc = unboxchar.getLoc(); |
3109 | mlir::Value ptrToBuffer = |
3110 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, tuple, 0); |
3111 | |
3112 | auto len = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, tuple, 1); |
3113 | mlir::Value lenAfterCast = integerCast(loc, rewriter, lenTy, len); |
3114 | |
3115 | rewriter.replaceOp(unboxchar, |
3116 | llvm::ArrayRef<mlir::Value>{ptrToBuffer, lenAfterCast}); |
3117 | return mlir::success(); |
3118 | } |
3119 | }; |
3120 | |
3121 | /// Lower `fir.unboxproc` operation. Unbox a procedure box value, yielding its |
3122 | /// components. |
3123 | /// TODO: Part of supporting Fortran 2003 procedure pointers. |
3124 | struct UnboxProcOpConversion : public fir::FIROpConversion<fir::UnboxProcOp> { |
3125 | using FIROpConversion::FIROpConversion; |
3126 | |
3127 | mlir::LogicalResult |
3128 | matchAndRewrite(fir::UnboxProcOp unboxproc, OpAdaptor adaptor, |
3129 | mlir::ConversionPatternRewriter &rewriter) const override { |
3130 | TODO(unboxproc.getLoc(), "fir.unboxproc codegen" ); |
3131 | return mlir::failure(); |
3132 | } |
3133 | }; |
3134 | |
3135 | /// convert to LLVM IR dialect `undef` |
3136 | struct UndefOpConversion : public fir::FIROpConversion<fir::UndefOp> { |
3137 | using FIROpConversion::FIROpConversion; |
3138 | |
3139 | mlir::LogicalResult |
3140 | matchAndRewrite(fir::UndefOp undef, OpAdaptor, |
3141 | mlir::ConversionPatternRewriter &rewriter) const override { |
3142 | rewriter.replaceOpWithNewOp<mlir::LLVM::UndefOp>( |
3143 | undef, convertType(undef.getType())); |
3144 | return mlir::success(); |
3145 | } |
3146 | }; |
3147 | |
3148 | struct ZeroOpConversion : public fir::FIROpConversion<fir::ZeroOp> { |
3149 | using FIROpConversion::FIROpConversion; |
3150 | |
3151 | mlir::LogicalResult |
3152 | matchAndRewrite(fir::ZeroOp zero, OpAdaptor, |
3153 | mlir::ConversionPatternRewriter &rewriter) const override { |
3154 | mlir::Type ty = convertType(zero.getType()); |
3155 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZeroOp>(zero, ty); |
3156 | return mlir::success(); |
3157 | } |
3158 | }; |
3159 | |
3160 | /// `fir.unreachable` --> `llvm.unreachable` |
3161 | struct UnreachableOpConversion |
3162 | : public fir::FIROpConversion<fir::UnreachableOp> { |
3163 | using FIROpConversion::FIROpConversion; |
3164 | |
3165 | mlir::LogicalResult |
3166 | matchAndRewrite(fir::UnreachableOp unreach, OpAdaptor adaptor, |
3167 | mlir::ConversionPatternRewriter &rewriter) const override { |
3168 | rewriter.replaceOpWithNewOp<mlir::LLVM::UnreachableOp>(unreach); |
3169 | return mlir::success(); |
3170 | } |
3171 | }; |
3172 | |
3173 | /// `fir.is_present` --> |
3174 | /// ``` |
3175 | /// %0 = llvm.mlir.constant(0 : i64) |
3176 | /// %1 = llvm.ptrtoint %0 |
3177 | /// %2 = llvm.icmp "ne" %1, %0 : i64 |
3178 | /// ``` |
3179 | struct IsPresentOpConversion : public fir::FIROpConversion<fir::IsPresentOp> { |
3180 | using FIROpConversion::FIROpConversion; |
3181 | |
3182 | mlir::LogicalResult |
3183 | matchAndRewrite(fir::IsPresentOp isPresent, OpAdaptor adaptor, |
3184 | mlir::ConversionPatternRewriter &rewriter) const override { |
3185 | mlir::Type idxTy = lowerTy().indexType(); |
3186 | mlir::Location loc = isPresent.getLoc(); |
3187 | auto ptr = adaptor.getOperands()[0]; |
3188 | |
3189 | if (isPresent.getVal().getType().isa<fir::BoxCharType>()) { |
3190 | [[maybe_unused]] auto structTy = |
3191 | ptr.getType().cast<mlir::LLVM::LLVMStructType>(); |
3192 | assert(!structTy.isOpaque() && !structTy.getBody().empty()); |
3193 | |
3194 | ptr = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, ptr, 0); |
3195 | } |
3196 | mlir::LLVM::ConstantOp c0 = |
3197 | genConstantIndex(isPresent.getLoc(), idxTy, rewriter, 0); |
3198 | auto addr = rewriter.create<mlir::LLVM::PtrToIntOp>(loc, idxTy, ptr); |
3199 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
3200 | isPresent, mlir::LLVM::ICmpPredicate::ne, addr, c0); |
3201 | |
3202 | return mlir::success(); |
3203 | } |
3204 | }; |
3205 | |
3206 | /// Create value signaling an absent optional argument in a call, e.g. |
3207 | /// `fir.absent !fir.ref<i64>` --> `llvm.mlir.zero : !llvm.ptr<i64>` |
3208 | struct AbsentOpConversion : public fir::FIROpConversion<fir::AbsentOp> { |
3209 | using FIROpConversion::FIROpConversion; |
3210 | |
3211 | mlir::LogicalResult |
3212 | matchAndRewrite(fir::AbsentOp absent, OpAdaptor, |
3213 | mlir::ConversionPatternRewriter &rewriter) const override { |
3214 | mlir::Type ty = convertType(absent.getType()); |
3215 | mlir::Location loc = absent.getLoc(); |
3216 | |
3217 | if (absent.getType().isa<fir::BoxCharType>()) { |
3218 | auto structTy = ty.cast<mlir::LLVM::LLVMStructType>(); |
3219 | assert(!structTy.isOpaque() && !structTy.getBody().empty()); |
3220 | auto undefStruct = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
3221 | auto nullField = |
3222 | rewriter.create<mlir::LLVM::ZeroOp>(loc, structTy.getBody()[0]); |
3223 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
3224 | absent, undefStruct, nullField, 0); |
3225 | } else { |
3226 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZeroOp>(absent, ty); |
3227 | } |
3228 | return mlir::success(); |
3229 | } |
3230 | }; |
3231 | |
3232 | // |
3233 | // Primitive operations on Complex types |
3234 | // |
3235 | |
3236 | template <typename OPTY> |
3237 | static inline mlir::LLVM::FastmathFlagsAttr getLLVMFMFAttr(OPTY op) { |
3238 | return mlir::LLVM::FastmathFlagsAttr::get( |
3239 | op.getContext(), |
3240 | mlir::arith::convertArithFastMathFlagsToLLVM(op.getFastmath())); |
3241 | } |
3242 | |
3243 | /// Generate inline code for complex addition/subtraction |
3244 | template <typename LLVMOP, typename OPTY> |
3245 | static mlir::LLVM::InsertValueOp |
3246 | complexSum(OPTY sumop, mlir::ValueRange opnds, |
3247 | mlir::ConversionPatternRewriter &rewriter, |
3248 | const fir::LLVMTypeConverter &lowering) { |
3249 | mlir::LLVM::FastmathFlagsAttr fmf = getLLVMFMFAttr(sumop); |
3250 | mlir::Value a = opnds[0]; |
3251 | mlir::Value b = opnds[1]; |
3252 | auto loc = sumop.getLoc(); |
3253 | mlir::Type eleTy = lowering.convertType(getComplexEleTy(sumop.getType())); |
3254 | mlir::Type ty = lowering.convertType(sumop.getType()); |
3255 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); |
3256 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); |
3257 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); |
3258 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); |
3259 | auto rx = rewriter.create<LLVMOP>(loc, eleTy, x0, x1, fmf); |
3260 | auto ry = rewriter.create<LLVMOP>(loc, eleTy, y0, y1, fmf); |
3261 | auto r0 = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
3262 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r0, rx, 0); |
3263 | return rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ry, 1); |
3264 | } |
3265 | } // namespace |
3266 | |
3267 | namespace { |
3268 | struct AddcOpConversion : public fir::FIROpConversion<fir::AddcOp> { |
3269 | using FIROpConversion::FIROpConversion; |
3270 | |
3271 | mlir::LogicalResult |
3272 | matchAndRewrite(fir::AddcOp addc, OpAdaptor adaptor, |
3273 | mlir::ConversionPatternRewriter &rewriter) const override { |
3274 | // given: (x + iy) + (x' + iy') |
3275 | // result: (x + x') + i(y + y') |
3276 | auto r = complexSum<mlir::LLVM::FAddOp>(addc, adaptor.getOperands(), |
3277 | rewriter, lowerTy()); |
3278 | rewriter.replaceOp(addc, r.getResult()); |
3279 | return mlir::success(); |
3280 | } |
3281 | }; |
3282 | |
3283 | struct SubcOpConversion : public fir::FIROpConversion<fir::SubcOp> { |
3284 | using FIROpConversion::FIROpConversion; |
3285 | |
3286 | mlir::LogicalResult |
3287 | matchAndRewrite(fir::SubcOp subc, OpAdaptor adaptor, |
3288 | mlir::ConversionPatternRewriter &rewriter) const override { |
3289 | // given: (x + iy) - (x' + iy') |
3290 | // result: (x - x') + i(y - y') |
3291 | auto r = complexSum<mlir::LLVM::FSubOp>(subc, adaptor.getOperands(), |
3292 | rewriter, lowerTy()); |
3293 | rewriter.replaceOp(subc, r.getResult()); |
3294 | return mlir::success(); |
3295 | } |
3296 | }; |
3297 | |
3298 | /// Inlined complex multiply |
3299 | struct MulcOpConversion : public fir::FIROpConversion<fir::MulcOp> { |
3300 | using FIROpConversion::FIROpConversion; |
3301 | |
3302 | mlir::LogicalResult |
3303 | matchAndRewrite(fir::MulcOp mulc, OpAdaptor adaptor, |
3304 | mlir::ConversionPatternRewriter &rewriter) const override { |
3305 | // TODO: Can we use a call to __muldc3 ? |
3306 | // given: (x + iy) * (x' + iy') |
3307 | // result: (xx'-yy')+i(xy'+yx') |
3308 | mlir::LLVM::FastmathFlagsAttr fmf = getLLVMFMFAttr(mulc); |
3309 | mlir::Value a = adaptor.getOperands()[0]; |
3310 | mlir::Value b = adaptor.getOperands()[1]; |
3311 | auto loc = mulc.getLoc(); |
3312 | mlir::Type eleTy = convertType(getComplexEleTy(mulc.getType())); |
3313 | mlir::Type ty = convertType(mulc.getType()); |
3314 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); |
3315 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); |
3316 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); |
3317 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); |
3318 | auto xx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, x1, fmf); |
3319 | auto yx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, x1, fmf); |
3320 | auto xy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, y1, fmf); |
3321 | auto ri = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, xy, yx, fmf); |
3322 | auto yy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, y1, fmf); |
3323 | auto rr = rewriter.create<mlir::LLVM::FSubOp>(loc, eleTy, xx, yy, fmf); |
3324 | auto ra = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
3325 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, ra, rr, 0); |
3326 | auto r0 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ri, 1); |
3327 | rewriter.replaceOp(mulc, r0.getResult()); |
3328 | return mlir::success(); |
3329 | } |
3330 | }; |
3331 | |
3332 | /// Inlined complex division |
3333 | struct DivcOpConversion : public fir::FIROpConversion<fir::DivcOp> { |
3334 | using FIROpConversion::FIROpConversion; |
3335 | |
3336 | mlir::LogicalResult |
3337 | matchAndRewrite(fir::DivcOp divc, OpAdaptor adaptor, |
3338 | mlir::ConversionPatternRewriter &rewriter) const override { |
3339 | // TODO: Can we use a call to __divdc3 instead? |
3340 | // Just generate inline code for now. |
3341 | // given: (x + iy) / (x' + iy') |
3342 | // result: ((xx'+yy')/d) + i((yx'-xy')/d) where d = x'x' + y'y' |
3343 | mlir::LLVM::FastmathFlagsAttr fmf = getLLVMFMFAttr(divc); |
3344 | mlir::Value a = adaptor.getOperands()[0]; |
3345 | mlir::Value b = adaptor.getOperands()[1]; |
3346 | auto loc = divc.getLoc(); |
3347 | mlir::Type eleTy = convertType(getComplexEleTy(divc.getType())); |
3348 | mlir::Type ty = convertType(divc.getType()); |
3349 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); |
3350 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); |
3351 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); |
3352 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); |
3353 | auto xx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, x1, fmf); |
3354 | auto x1x1 = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x1, x1, fmf); |
3355 | auto yx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, x1, fmf); |
3356 | auto xy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, y1, fmf); |
3357 | auto yy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, y1, fmf); |
3358 | auto y1y1 = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y1, y1, fmf); |
3359 | auto d = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, x1x1, y1y1, fmf); |
3360 | auto rrn = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, xx, yy, fmf); |
3361 | auto rin = rewriter.create<mlir::LLVM::FSubOp>(loc, eleTy, yx, xy, fmf); |
3362 | auto rr = rewriter.create<mlir::LLVM::FDivOp>(loc, eleTy, rrn, d, fmf); |
3363 | auto ri = rewriter.create<mlir::LLVM::FDivOp>(loc, eleTy, rin, d, fmf); |
3364 | auto ra = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); |
3365 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, ra, rr, 0); |
3366 | auto r0 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ri, 1); |
3367 | rewriter.replaceOp(divc, r0.getResult()); |
3368 | return mlir::success(); |
3369 | } |
3370 | }; |
3371 | |
3372 | /// Inlined complex negation |
3373 | struct NegcOpConversion : public fir::FIROpConversion<fir::NegcOp> { |
3374 | using FIROpConversion::FIROpConversion; |
3375 | |
3376 | mlir::LogicalResult |
3377 | matchAndRewrite(fir::NegcOp neg, OpAdaptor adaptor, |
3378 | mlir::ConversionPatternRewriter &rewriter) const override { |
3379 | // given: -(x + iy) |
3380 | // result: -x - iy |
3381 | auto eleTy = convertType(getComplexEleTy(neg.getType())); |
3382 | auto loc = neg.getLoc(); |
3383 | mlir::Value o0 = adaptor.getOperands()[0]; |
3384 | auto rp = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, o0, 0); |
3385 | auto ip = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, o0, 1); |
3386 | auto nrp = rewriter.create<mlir::LLVM::FNegOp>(loc, eleTy, rp); |
3387 | auto nip = rewriter.create<mlir::LLVM::FNegOp>(loc, eleTy, ip); |
3388 | auto r = rewriter.create<mlir::LLVM::InsertValueOp>(loc, o0, nrp, 0); |
3389 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(neg, r, nip, 1); |
3390 | return mlir::success(); |
3391 | } |
3392 | }; |
3393 | |
3394 | struct BoxOffsetOpConversion : public fir::FIROpConversion<fir::BoxOffsetOp> { |
3395 | using FIROpConversion::FIROpConversion; |
3396 | |
3397 | mlir::LogicalResult |
3398 | matchAndRewrite(fir::BoxOffsetOp boxOffset, OpAdaptor adaptor, |
3399 | mlir::ConversionPatternRewriter &rewriter) const override { |
3400 | |
3401 | mlir::Type pty = ::getLlvmPtrType(boxOffset.getContext()); |
3402 | mlir::Type boxType = fir::unwrapRefType(boxOffset.getBoxRef().getType()); |
3403 | mlir::Type llvmBoxTy = |
3404 | lowerTy().convertBoxTypeAsStruct(mlir::cast<fir::BaseBoxType>(boxType)); |
3405 | int fieldId = boxOffset.getField() == fir::BoxFieldAttr::derived_type |
3406 | ? getTypeDescFieldId(boxType) |
3407 | : kAddrPosInBox; |
3408 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
3409 | boxOffset, pty, llvmBoxTy, adaptor.getBoxRef(), |
3410 | llvm::ArrayRef<mlir::LLVM::GEPArg>{0, fieldId}); |
3411 | return mlir::success(); |
3412 | } |
3413 | }; |
3414 | |
3415 | /// Conversion pattern for operation that must be dead. The information in these |
3416 | /// operations is used by other operation. At this point they should not have |
3417 | /// anymore uses. |
3418 | /// These operations are normally dead after the pre-codegen pass. |
3419 | template <typename FromOp> |
3420 | struct MustBeDeadConversion : public fir::FIROpConversion<FromOp> { |
3421 | explicit MustBeDeadConversion(const fir::LLVMTypeConverter &lowering, |
3422 | const fir::FIRToLLVMPassOptions &options) |
3423 | : fir::FIROpConversion<FromOp>(lowering, options) {} |
3424 | using OpAdaptor = typename FromOp::Adaptor; |
3425 | |
3426 | mlir::LogicalResult |
3427 | matchAndRewrite(FromOp op, OpAdaptor adaptor, |
3428 | mlir::ConversionPatternRewriter &rewriter) const final { |
3429 | if (!op->getUses().empty()) |
3430 | return rewriter.notifyMatchFailure(op, "op must be dead" ); |
3431 | rewriter.eraseOp(op); |
3432 | return mlir::success(); |
3433 | } |
3434 | }; |
3435 | |
3436 | struct UnrealizedConversionCastOpConversion |
3437 | : public fir::FIROpConversion<mlir::UnrealizedConversionCastOp> { |
3438 | using FIROpConversion::FIROpConversion; |
3439 | |
3440 | mlir::LogicalResult |
3441 | matchAndRewrite(mlir::UnrealizedConversionCastOp op, OpAdaptor adaptor, |
3442 | mlir::ConversionPatternRewriter &rewriter) const override { |
3443 | assert(op.getOutputs().getTypes().size() == 1 && "expect a single type" ); |
3444 | mlir::Type convertedType = convertType(op.getOutputs().getTypes()[0]); |
3445 | if (convertedType == adaptor.getInputs().getTypes()[0]) { |
3446 | rewriter.replaceOp(op, adaptor.getInputs()); |
3447 | return mlir::success(); |
3448 | } |
3449 | |
3450 | convertedType = adaptor.getInputs().getTypes()[0]; |
3451 | if (convertedType == op.getOutputs().getType()[0]) { |
3452 | rewriter.replaceOp(op, adaptor.getInputs()); |
3453 | return mlir::success(); |
3454 | } |
3455 | return mlir::failure(); |
3456 | } |
3457 | }; |
3458 | |
3459 | struct ShapeOpConversion : public MustBeDeadConversion<fir::ShapeOp> { |
3460 | using MustBeDeadConversion::MustBeDeadConversion; |
3461 | }; |
3462 | |
3463 | struct ShapeShiftOpConversion : public MustBeDeadConversion<fir::ShapeShiftOp> { |
3464 | using MustBeDeadConversion::MustBeDeadConversion; |
3465 | }; |
3466 | |
3467 | struct ShiftOpConversion : public MustBeDeadConversion<fir::ShiftOp> { |
3468 | using MustBeDeadConversion::MustBeDeadConversion; |
3469 | }; |
3470 | |
3471 | struct SliceOpConversion : public MustBeDeadConversion<fir::SliceOp> { |
3472 | using MustBeDeadConversion::MustBeDeadConversion; |
3473 | }; |
3474 | |
3475 | } // namespace |
3476 | |
3477 | namespace { |
3478 | class RenameMSVCLibmCallees |
3479 | : public mlir::OpRewritePattern<mlir::LLVM::CallOp> { |
3480 | public: |
3481 | using OpRewritePattern::OpRewritePattern; |
3482 | |
3483 | mlir::LogicalResult |
3484 | matchAndRewrite(mlir::LLVM::CallOp op, |
3485 | mlir::PatternRewriter &rewriter) const override { |
3486 | rewriter.startOpModification(op); |
3487 | auto callee = op.getCallee(); |
3488 | if (callee) |
3489 | if (callee->equals("hypotf" )) |
3490 | op.setCalleeAttr(mlir::SymbolRefAttr::get(op.getContext(), "_hypotf" )); |
3491 | |
3492 | rewriter.finalizeOpModification(op); |
3493 | return mlir::success(); |
3494 | } |
3495 | }; |
3496 | |
3497 | class RenameMSVCLibmFuncs |
3498 | : public mlir::OpRewritePattern<mlir::LLVM::LLVMFuncOp> { |
3499 | public: |
3500 | using OpRewritePattern::OpRewritePattern; |
3501 | |
3502 | mlir::LogicalResult |
3503 | matchAndRewrite(mlir::LLVM::LLVMFuncOp op, |
3504 | mlir::PatternRewriter &rewriter) const override { |
3505 | rewriter.startOpModification(op); |
3506 | if (op.getSymName().equals("hypotf" )) |
3507 | op.setSymNameAttr(rewriter.getStringAttr("_hypotf" )); |
3508 | rewriter.finalizeOpModification(op); |
3509 | return mlir::success(); |
3510 | } |
3511 | }; |
3512 | } // namespace |
3513 | |
3514 | namespace { |
3515 | /// Convert FIR dialect to LLVM dialect |
3516 | /// |
3517 | /// This pass lowers all FIR dialect operations to LLVM IR dialect. An |
3518 | /// MLIR pass is used to lower residual Std dialect to LLVM IR dialect. |
3519 | class FIRToLLVMLowering |
3520 | : public fir::impl::FIRToLLVMLoweringBase<FIRToLLVMLowering> { |
3521 | public: |
3522 | FIRToLLVMLowering() = default; |
3523 | FIRToLLVMLowering(fir::FIRToLLVMPassOptions options) : options{options} {} |
3524 | mlir::ModuleOp getModule() { return getOperation(); } |
3525 | |
3526 | void runOnOperation() override final { |
3527 | auto mod = getModule(); |
3528 | if (!forcedTargetTriple.empty()) |
3529 | fir::setTargetTriple(mod, forcedTargetTriple); |
3530 | |
3531 | if (!forcedDataLayout.empty()) { |
3532 | llvm::DataLayout dl(forcedDataLayout); |
3533 | fir::support::setMLIRDataLayout(mod, dl); |
3534 | } |
3535 | |
3536 | if (!forcedTargetCPU.empty()) |
3537 | fir::setTargetCPU(mod, forcedTargetCPU); |
3538 | |
3539 | if (!forcedTargetFeatures.empty()) |
3540 | fir::setTargetFeatures(mod, forcedTargetFeatures); |
3541 | |
3542 | // Run dynamic pass pipeline for converting Math dialect |
3543 | // operations into other dialects (llvm, func, etc.). |
3544 | // Some conversions of Math operations cannot be done |
3545 | // by just using conversion patterns. This is true for |
3546 | // conversions that affect the ModuleOp, e.g. create new |
3547 | // function operations in it. We have to run such conversions |
3548 | // as passes here. |
3549 | mlir::OpPassManager mathConvertionPM("builtin.module" ); |
3550 | |
3551 | // Convert math::FPowI operations to inline implementation |
3552 | // only if the exponent's width is greater than 32, otherwise, |
3553 | // it will be lowered to LLVM intrinsic operation by a later conversion. |
3554 | mlir::ConvertMathToFuncsOptions mathToFuncsOptions{}; |
3555 | mathToFuncsOptions.minWidthOfFPowIExponent = 33; |
3556 | mathConvertionPM.addPass( |
3557 | mlir::createConvertMathToFuncs(mathToFuncsOptions)); |
3558 | mathConvertionPM.addPass(mlir::createConvertComplexToStandardPass()); |
3559 | // Convert Math dialect operations into LLVM dialect operations. |
3560 | // There is no way to prefer MathToLLVM patterns over MathToLibm |
3561 | // patterns (applied below), so we have to run MathToLLVM conversion here. |
3562 | mathConvertionPM.addNestedPass<mlir::func::FuncOp>( |
3563 | mlir::createConvertMathToLLVMPass()); |
3564 | if (mlir::failed(runPipeline(mathConvertionPM, mod))) |
3565 | return signalPassFailure(); |
3566 | |
3567 | std::optional<mlir::DataLayout> dl = |
3568 | fir::support::getOrSetDataLayout(mod, /*allowDefaultLayout=*/true); |
3569 | if (!dl) { |
3570 | mlir::emitError(mod.getLoc(), |
3571 | "module operation must carry a data layout attribute " |
3572 | "to generate llvm IR from FIR" ); |
3573 | signalPassFailure(); |
3574 | return; |
3575 | } |
3576 | |
3577 | auto *context = getModule().getContext(); |
3578 | fir::LLVMTypeConverter typeConverter{getModule(), |
3579 | options.applyTBAA || applyTBAA, |
3580 | options.forceUnifiedTBAATree, *dl}; |
3581 | mlir::RewritePatternSet pattern(context); |
3582 | fir::populateFIRToLLVMConversionPatterns(typeConverter, pattern, options); |
3583 | mlir::populateFuncToLLVMConversionPatterns(typeConverter, pattern); |
3584 | mlir::populateOpenMPToLLVMConversionPatterns(typeConverter, pattern); |
3585 | mlir::arith::populateArithToLLVMConversionPatterns(typeConverter, pattern); |
3586 | mlir::cf::populateControlFlowToLLVMConversionPatterns(typeConverter, |
3587 | pattern); |
3588 | // Math operations that have not been converted yet must be converted |
3589 | // to Libm. |
3590 | mlir::populateMathToLibmConversionPatterns(pattern); |
3591 | mlir::populateComplexToLLVMConversionPatterns(typeConverter, pattern); |
3592 | mlir::populateVectorToLLVMConversionPatterns(typeConverter, pattern); |
3593 | |
3594 | // Flang specific overloads for OpenMP operations, to allow for special |
3595 | // handling of things like Box types. |
3596 | fir::populateOpenMPFIRToLLVMConversionPatterns(typeConverter, pattern); |
3597 | |
3598 | mlir::ConversionTarget target{*context}; |
3599 | target.addLegalDialect<mlir::LLVM::LLVMDialect>(); |
3600 | // The OpenMP dialect is legal for Operations without regions, for those |
3601 | // which contains regions it is legal if the region contains only the |
3602 | // LLVM dialect. Add OpenMP dialect as a legal dialect for conversion and |
3603 | // legalize conversion of OpenMP operations without regions. |
3604 | mlir::configureOpenMPToLLVMConversionLegality(target, typeConverter); |
3605 | target.addLegalDialect<mlir::omp::OpenMPDialect>(); |
3606 | target.addLegalDialect<mlir::acc::OpenACCDialect>(); |
3607 | |
3608 | // required NOPs for applying a full conversion |
3609 | target.addLegalOp<mlir::ModuleOp>(); |
3610 | |
3611 | // If we're on Windows, we might need to rename some libm calls. |
3612 | bool isMSVC = fir::getTargetTriple(mod).isOSMSVCRT(); |
3613 | if (isMSVC) { |
3614 | pattern.insert<RenameMSVCLibmCallees, RenameMSVCLibmFuncs>(context); |
3615 | |
3616 | target.addDynamicallyLegalOp<mlir::LLVM::CallOp>( |
3617 | [](mlir::LLVM::CallOp op) { |
3618 | auto callee = op.getCallee(); |
3619 | if (!callee) |
3620 | return true; |
3621 | return !callee->equals("hypotf" ); |
3622 | }); |
3623 | target.addDynamicallyLegalOp<mlir::LLVM::LLVMFuncOp>( |
3624 | [](mlir::LLVM::LLVMFuncOp op) { |
3625 | return !op.getSymName().equals("hypotf" ); |
3626 | }); |
3627 | } |
3628 | |
3629 | // apply the patterns |
3630 | if (mlir::failed(mlir::applyFullConversion(getModule(), target, |
3631 | std::move(pattern)))) { |
3632 | signalPassFailure(); |
3633 | } |
3634 | |
3635 | // Run pass to add comdats to functions that have weak linkage on relevant platforms |
3636 | if (fir::getTargetTriple(mod).supportsCOMDAT()) { |
3637 | mlir::OpPassManager comdatPM("builtin.module" ); |
3638 | comdatPM.addPass(mlir::LLVM::createLLVMAddComdats()); |
3639 | if (mlir::failed(runPipeline(comdatPM, mod))) |
3640 | return signalPassFailure(); |
3641 | } |
3642 | } |
3643 | |
3644 | private: |
3645 | fir::FIRToLLVMPassOptions options; |
3646 | }; |
3647 | |
3648 | /// Lower from LLVM IR dialect to proper LLVM-IR and dump the module |
3649 | struct LLVMIRLoweringPass |
3650 | : public mlir::PassWrapper<LLVMIRLoweringPass, |
3651 | mlir::OperationPass<mlir::ModuleOp>> { |
3652 | MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(LLVMIRLoweringPass) |
3653 | |
3654 | LLVMIRLoweringPass(llvm::raw_ostream &output, fir::LLVMIRLoweringPrinter p) |
3655 | : output{output}, printer{p} {} |
3656 | |
3657 | mlir::ModuleOp getModule() { return getOperation(); } |
3658 | |
3659 | void runOnOperation() override final { |
3660 | auto *ctx = getModule().getContext(); |
3661 | auto optName = getModule().getName(); |
3662 | llvm::LLVMContext llvmCtx; |
3663 | if (auto llvmModule = mlir::translateModuleToLLVMIR( |
3664 | getModule(), llvmCtx, optName ? *optName : "FIRModule" )) { |
3665 | printer(*llvmModule, output); |
3666 | return; |
3667 | } |
3668 | |
3669 | mlir::emitError(mlir::UnknownLoc::get(ctx), "could not emit LLVM-IR\n" ); |
3670 | signalPassFailure(); |
3671 | } |
3672 | |
3673 | private: |
3674 | llvm::raw_ostream &output; |
3675 | fir::LLVMIRLoweringPrinter printer; |
3676 | }; |
3677 | |
3678 | } // namespace |
3679 | |
3680 | std::unique_ptr<mlir::Pass> fir::createFIRToLLVMPass() { |
3681 | return std::make_unique<FIRToLLVMLowering>(); |
3682 | } |
3683 | |
3684 | std::unique_ptr<mlir::Pass> |
3685 | fir::createFIRToLLVMPass(fir::FIRToLLVMPassOptions options) { |
3686 | return std::make_unique<FIRToLLVMLowering>(options); |
3687 | } |
3688 | |
3689 | std::unique_ptr<mlir::Pass> |
3690 | fir::createLLVMDialectToLLVMPass(llvm::raw_ostream &output, |
3691 | fir::LLVMIRLoweringPrinter printer) { |
3692 | return std::make_unique<LLVMIRLoweringPass>(output, printer); |
3693 | } |
3694 | |
3695 | void fir::populateFIRToLLVMConversionPatterns( |
3696 | fir::LLVMTypeConverter &converter, mlir::RewritePatternSet &patterns, |
3697 | fir::FIRToLLVMPassOptions &options) { |
3698 | patterns.insert< |
3699 | AbsentOpConversion, AddcOpConversion, AddrOfOpConversion, |
3700 | AllocaOpConversion, AllocMemOpConversion, BoxAddrOpConversion, |
3701 | BoxCharLenOpConversion, BoxDimsOpConversion, BoxEleSizeOpConversion, |
3702 | BoxIsAllocOpConversion, BoxIsArrayOpConversion, BoxIsPtrOpConversion, |
3703 | BoxOffsetOpConversion, BoxProcHostOpConversion, BoxRankOpConversion, |
3704 | BoxTypeCodeOpConversion, BoxTypeDescOpConversion, CallOpConversion, |
3705 | CmpcOpConversion, ConstcOpConversion, ConvertOpConversion, |
3706 | CoordinateOpConversion, DTEntryOpConversion, DivcOpConversion, |
3707 | EmboxOpConversion, EmboxCharOpConversion, EmboxProcOpConversion, |
3708 | ExtractValueOpConversion, FieldIndexOpConversion, FirEndOpConversion, |
3709 | FreeMemOpConversion, GlobalLenOpConversion, GlobalOpConversion, |
3710 | HasValueOpConversion, InsertOnRangeOpConversion, InsertValueOpConversion, |
3711 | IsPresentOpConversion, LenParamIndexOpConversion, LoadOpConversion, |
3712 | MulcOpConversion, NegcOpConversion, NoReassocOpConversion, |
3713 | SelectCaseOpConversion, SelectOpConversion, SelectRankOpConversion, |
3714 | SelectTypeOpConversion, ShapeOpConversion, ShapeShiftOpConversion, |
3715 | ShiftOpConversion, SliceOpConversion, StoreOpConversion, |
3716 | StringLitOpConversion, SubcOpConversion, TypeDescOpConversion, |
3717 | TypeInfoOpConversion, UnboxCharOpConversion, UnboxProcOpConversion, |
3718 | UndefOpConversion, UnreachableOpConversion, |
3719 | UnrealizedConversionCastOpConversion, XArrayCoorOpConversion, |
3720 | XEmboxOpConversion, XReboxOpConversion, ZeroOpConversion>(converter, |
3721 | options); |
3722 | } |
3723 | |