| File: | build/source/llvm/../mlir/include/mlir/IR/Builders.h |
| Warning: | line 490, column 5 4th function call argument is an uninitialized value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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/ISO_Fortran_binding.h" | ||||||
| 17 | #include "flang/Optimizer/Dialect/FIRAttr.h" | ||||||
| 18 | #include "flang/Optimizer/Dialect/FIROps.h" | ||||||
| 19 | #include "flang/Optimizer/Dialect/FIRType.h" | ||||||
| 20 | #include "flang/Optimizer/Support/InternalNames.h" | ||||||
| 21 | #include "flang/Optimizer/Support/TypeCode.h" | ||||||
| 22 | #include "flang/Optimizer/Support/Utils.h" | ||||||
| 23 | #include "flang/Semantics/runtime-type-info.h" | ||||||
| 24 | #include "mlir/Conversion/ArithCommon/AttrToLLVMConverter.h" | ||||||
| 25 | #include "mlir/Conversion/ArithToLLVM/ArithToLLVM.h" | ||||||
| 26 | #include "mlir/Conversion/ComplexToLLVM/ComplexToLLVM.h" | ||||||
| 27 | #include "mlir/Conversion/ComplexToStandard/ComplexToStandard.h" | ||||||
| 28 | #include "mlir/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.h" | ||||||
| 29 | #include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVM.h" | ||||||
| 30 | #include "mlir/Conversion/LLVMCommon/Pattern.h" | ||||||
| 31 | #include "mlir/Conversion/MathToFuncs/MathToFuncs.h" | ||||||
| 32 | #include "mlir/Conversion/MathToLLVM/MathToLLVM.h" | ||||||
| 33 | #include "mlir/Conversion/MathToLibm/MathToLibm.h" | ||||||
| 34 | #include "mlir/Conversion/OpenACCToLLVM/ConvertOpenACCToLLVM.h" | ||||||
| 35 | #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h" | ||||||
| 36 | #include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h" | ||||||
| 37 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" | ||||||
| 38 | #include "mlir/Dialect/OpenACC/OpenACC.h" | ||||||
| 39 | #include "mlir/Dialect/OpenMP/OpenMPDialect.h" | ||||||
| 40 | #include "mlir/IR/BuiltinTypes.h" | ||||||
| 41 | #include "mlir/IR/Matchers.h" | ||||||
| 42 | #include "mlir/Pass/Pass.h" | ||||||
| 43 | #include "mlir/Pass/PassManager.h" | ||||||
| 44 | #include "mlir/Target/LLVMIR/ModuleTranslation.h" | ||||||
| 45 | #include "llvm/ADT/ArrayRef.h" | ||||||
| 46 | #include "llvm/ADT/TypeSwitch.h" | ||||||
| 47 | |||||||
| 48 | namespace fir { | ||||||
| 49 | #define GEN_PASS_DEF_FIRTOLLVMLOWERING | ||||||
| 50 | #include "flang/Optimizer/CodeGen/CGPasses.h.inc" | ||||||
| 51 | } // namespace fir | ||||||
| 52 | |||||||
| 53 | #define DEBUG_TYPE"flang-codegen" "flang-codegen" | ||||||
| 54 | |||||||
| 55 | // fir::LLVMTypeConverter for converting to LLVM IR dialect types. | ||||||
| 56 | #include "flang/Optimizer/CodeGen/TypeConverter.h" | ||||||
| 57 | |||||||
| 58 | // TODO: This should really be recovered from the specified target. | ||||||
| 59 | static constexpr unsigned defaultAlign = 8; | ||||||
| 60 | |||||||
| 61 | /// `fir.box` attribute values as defined for CFI_attribute_t in | ||||||
| 62 | /// flang/ISO_Fortran_binding.h. | ||||||
| 63 | static constexpr unsigned kAttrPointer = CFI_attribute_pointer1; | ||||||
| 64 | static constexpr unsigned kAttrAllocatable = CFI_attribute_allocatable2; | ||||||
| 65 | |||||||
| 66 | static inline mlir::Type getVoidPtrType(mlir::MLIRContext *context) { | ||||||
| 67 | return mlir::LLVM::LLVMPointerType::get(mlir::IntegerType::get(context, 8)); | ||||||
| 68 | } | ||||||
| 69 | |||||||
| 70 | static mlir::LLVM::ConstantOp | ||||||
| 71 | genConstantIndex(mlir::Location loc, mlir::Type ity, | ||||||
| 72 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 73 | std::int64_t offset) { | ||||||
| 74 | auto cattr = rewriter.getI64IntegerAttr(offset); | ||||||
| 75 | return rewriter.create<mlir::LLVM::ConstantOp>(loc, ity, cattr); | ||||||
| 76 | } | ||||||
| 77 | |||||||
| 78 | static mlir::Block *createBlock(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 79 | mlir::Block *insertBefore) { | ||||||
| 80 | assert(insertBefore && "expected valid insertion block")(static_cast <bool> (insertBefore && "expected valid insertion block" ) ? void (0) : __assert_fail ("insertBefore && \"expected valid insertion block\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 80, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 81 | return rewriter.createBlock(insertBefore->getParent(), | ||||||
| 82 | mlir::Region::iterator(insertBefore)); | ||||||
| 83 | } | ||||||
| 84 | |||||||
| 85 | /// Extract constant from a value if it is a result of one of the | ||||||
| 86 | /// ConstantOp operations, otherwise, return std::nullopt. | ||||||
| 87 | static std::optional<int64_t> getIfConstantIntValue(mlir::Value val) { | ||||||
| 88 | if (!val || !val.dyn_cast<mlir::OpResult>()) | ||||||
| 89 | return {}; | ||||||
| 90 | |||||||
| 91 | mlir::Operation *defop = val.getDefiningOp(); | ||||||
| 92 | |||||||
| 93 | if (auto constOp = mlir::dyn_cast<mlir::arith::ConstantIntOp>(defop)) | ||||||
| 94 | return constOp.value(); | ||||||
| 95 | if (auto llConstOp = mlir::dyn_cast<mlir::LLVM::ConstantOp>(defop)) | ||||||
| 96 | if (auto attr = llConstOp.getValue().dyn_cast<mlir::IntegerAttr>()) | ||||||
| 97 | return attr.getValue().getSExtValue(); | ||||||
| 98 | |||||||
| 99 | return {}; | ||||||
| 100 | } | ||||||
| 101 | |||||||
| 102 | /// Extract constant from a value that must be the result of one of the | ||||||
| 103 | /// ConstantOp operations. | ||||||
| 104 | static int64_t getConstantIntValue(mlir::Value val) { | ||||||
| 105 | if (auto constVal = getIfConstantIntValue(val)) | ||||||
| 106 | return *constVal; | ||||||
| 107 | fir::emitFatalError(val.getLoc(), "must be a constant"); | ||||||
| 108 | } | ||||||
| 109 | |||||||
| 110 | static unsigned getTypeDescFieldId(mlir::Type ty) { | ||||||
| 111 | auto isArray = fir::dyn_cast_ptrOrBoxEleTy(ty).isa<fir::SequenceType>(); | ||||||
| 112 | return isArray ? kOptTypePtrPosInBox : kDimsPosInBox; | ||||||
| 113 | } | ||||||
| 114 | |||||||
| 115 | namespace { | ||||||
| 116 | /// FIR conversion pattern template | ||||||
| 117 | template <typename FromOp> | ||||||
| 118 | class FIROpConversion : public mlir::ConvertOpToLLVMPattern<FromOp> { | ||||||
| 119 | public: | ||||||
| 120 | explicit FIROpConversion(fir::LLVMTypeConverter &lowering, | ||||||
| 121 | const fir::FIRToLLVMPassOptions &options) | ||||||
| 122 | : mlir::ConvertOpToLLVMPattern<FromOp>(lowering), options(options) {} | ||||||
| 123 | |||||||
| 124 | protected: | ||||||
| 125 | mlir::Type convertType(mlir::Type ty) const { | ||||||
| 126 | return lowerTy().convertType(ty); | ||||||
| 127 | } | ||||||
| 128 | mlir::Type voidPtrTy() const { return getVoidPtrType(); } | ||||||
| 129 | |||||||
| 130 | mlir::Type getVoidPtrType() const { | ||||||
| 131 | return mlir::LLVM::LLVMPointerType::get( | ||||||
| 132 | mlir::IntegerType::get(&lowerTy().getContext(), 8)); | ||||||
| 133 | } | ||||||
| 134 | |||||||
| 135 | mlir::LLVM::ConstantOp | ||||||
| 136 | genI32Constant(mlir::Location loc, mlir::ConversionPatternRewriter &rewriter, | ||||||
| 137 | int value) const { | ||||||
| 138 | mlir::Type i32Ty = rewriter.getI32Type(); | ||||||
| 139 | mlir::IntegerAttr attr = rewriter.getI32IntegerAttr(value); | ||||||
| 140 | return rewriter.create<mlir::LLVM::ConstantOp>(loc, i32Ty, attr); | ||||||
| 141 | } | ||||||
| 142 | |||||||
| 143 | mlir::LLVM::ConstantOp | ||||||
| 144 | genConstantOffset(mlir::Location loc, | ||||||
| 145 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 146 | int offset) const { | ||||||
| 147 | mlir::Type ity = lowerTy().offsetType(); | ||||||
| 148 | mlir::IntegerAttr cattr = rewriter.getI32IntegerAttr(offset); | ||||||
| 149 | return rewriter.create<mlir::LLVM::ConstantOp>(loc, ity, cattr); | ||||||
| 150 | } | ||||||
| 151 | |||||||
| 152 | /// Perform an extension or truncation as needed on an integer value. Lowering | ||||||
| 153 | /// to the specific target may involve some sign-extending or truncation of | ||||||
| 154 | /// values, particularly to fit them from abstract box types to the | ||||||
| 155 | /// appropriate reified structures. | ||||||
| 156 | mlir::Value integerCast(mlir::Location loc, | ||||||
| 157 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 158 | mlir::Type ty, mlir::Value val) const { | ||||||
| 159 | auto valTy = val.getType(); | ||||||
| 160 | // If the value was not yet lowered, lower its type so that it can | ||||||
| 161 | // be used in getPrimitiveTypeSizeInBits. | ||||||
| 162 | if (!valTy.isa<mlir::IntegerType>()) | ||||||
| 163 | valTy = convertType(valTy); | ||||||
| 164 | auto toSize = mlir::LLVM::getPrimitiveTypeSizeInBits(ty); | ||||||
| 165 | auto fromSize = mlir::LLVM::getPrimitiveTypeSizeInBits(valTy); | ||||||
| 166 | if (toSize < fromSize) | ||||||
| 167 | return rewriter.create<mlir::LLVM::TruncOp>(loc, ty, val); | ||||||
| 168 | if (toSize > fromSize) | ||||||
| 169 | return rewriter.create<mlir::LLVM::SExtOp>(loc, ty, val); | ||||||
| 170 | return val; | ||||||
| 171 | } | ||||||
| 172 | |||||||
| 173 | /// Construct code sequence to extract the specific value from a `fir.box`. | ||||||
| 174 | mlir::Value getValueFromBox(mlir::Location loc, mlir::Type boxTy, | ||||||
| 175 | mlir::Value box, mlir::Type resultTy, | ||||||
| 176 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 177 | int boxValue) const { | ||||||
| 178 | if (box.getType().isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 179 | auto pty = mlir::LLVM::LLVMPointerType::get(resultTy); | ||||||
| 180 | auto p = rewriter.create<mlir::LLVM::GEPOp>( | ||||||
| 181 | loc, pty, box, llvm::ArrayRef<mlir::LLVM::GEPArg>{0, boxValue}); | ||||||
| 182 | auto loadOp = rewriter.create<mlir::LLVM::LoadOp>(loc, resultTy, p); | ||||||
| 183 | attachTBAATag(loadOp, boxTy, nullptr, p); | ||||||
| 184 | return loadOp; | ||||||
| 185 | } | ||||||
| 186 | return rewriter.create<mlir::LLVM::ExtractValueOp>(loc, box, boxValue); | ||||||
| 187 | } | ||||||
| 188 | |||||||
| 189 | /// Method to construct code sequence to get the triple for dimension `dim` | ||||||
| 190 | /// from a box. | ||||||
| 191 | llvm::SmallVector<mlir::Value, 3> | ||||||
| 192 | getDimsFromBox(mlir::Location loc, llvm::ArrayRef<mlir::Type> retTys, | ||||||
| 193 | mlir::Type boxTy, mlir::Value box, mlir::Value dim, | ||||||
| 194 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 195 | mlir::Value l0 = | ||||||
| 196 | loadDimFieldFromBox(loc, boxTy, box, dim, 0, retTys[0], rewriter); | ||||||
| 197 | mlir::Value l1 = | ||||||
| 198 | loadDimFieldFromBox(loc, boxTy, box, dim, 1, retTys[1], rewriter); | ||||||
| 199 | mlir::Value l2 = | ||||||
| 200 | loadDimFieldFromBox(loc, boxTy, box, dim, 2, retTys[2], rewriter); | ||||||
| 201 | return {l0, l1, l2}; | ||||||
| 202 | } | ||||||
| 203 | |||||||
| 204 | llvm::SmallVector<mlir::Value, 3> | ||||||
| 205 | getDimsFromBox(mlir::Location loc, llvm::ArrayRef<mlir::Type> retTys, | ||||||
| 206 | mlir::Type boxTy, mlir::Value box, int dim, | ||||||
| 207 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 208 | mlir::Value l0 = | ||||||
| 209 | getDimFieldFromBox(loc, boxTy, box, dim, 0, retTys[0], rewriter); | ||||||
| 210 | mlir::Value l1 = | ||||||
| 211 | getDimFieldFromBox(loc, boxTy, box, dim, 1, retTys[1], rewriter); | ||||||
| 212 | mlir::Value l2 = | ||||||
| 213 | getDimFieldFromBox(loc, boxTy, box, dim, 2, retTys[2], rewriter); | ||||||
| 214 | return {l0, l1, l2}; | ||||||
| 215 | } | ||||||
| 216 | |||||||
| 217 | mlir::Value | ||||||
| 218 | loadDimFieldFromBox(mlir::Location loc, mlir::Type boxTy, mlir::Value box, | ||||||
| 219 | mlir::Value dim, int off, mlir::Type ty, | ||||||
| 220 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 221 | assert(box.getType().isa<mlir::LLVM::LLVMPointerType>() &&(static_cast <bool> (box.getType().isa<mlir::LLVM::LLVMPointerType >() && "descriptor inquiry with runtime dim can only be done on descriptor " "in memory") ? void (0) : __assert_fail ("box.getType().isa<mlir::LLVM::LLVMPointerType>() && \"descriptor inquiry with runtime dim can only be done on descriptor \" \"in memory\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 223, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 222 | "descriptor inquiry with runtime dim can only be done on descriptor "(static_cast <bool> (box.getType().isa<mlir::LLVM::LLVMPointerType >() && "descriptor inquiry with runtime dim can only be done on descriptor " "in memory") ? void (0) : __assert_fail ("box.getType().isa<mlir::LLVM::LLVMPointerType>() && \"descriptor inquiry with runtime dim can only be done on descriptor \" \"in memory\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 223, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 223 | "in memory")(static_cast <bool> (box.getType().isa<mlir::LLVM::LLVMPointerType >() && "descriptor inquiry with runtime dim can only be done on descriptor " "in memory") ? void (0) : __assert_fail ("box.getType().isa<mlir::LLVM::LLVMPointerType>() && \"descriptor inquiry with runtime dim can only be done on descriptor \" \"in memory\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 223, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 224 | auto pty = mlir::LLVM::LLVMPointerType::get(ty); | ||||||
| 225 | mlir::LLVM::GEPOp p = genGEP(loc, pty, rewriter, box, 0, | ||||||
| 226 | static_cast<int>(kDimsPosInBox), dim, off); | ||||||
| 227 | auto loadOp = rewriter.create<mlir::LLVM::LoadOp>(loc, ty, p); | ||||||
| 228 | attachTBAATag(loadOp, boxTy, nullptr, p); | ||||||
| 229 | return loadOp; | ||||||
| 230 | } | ||||||
| 231 | |||||||
| 232 | mlir::Value | ||||||
| 233 | getDimFieldFromBox(mlir::Location loc, mlir::Type boxTy, mlir::Value box, | ||||||
| 234 | int dim, int off, mlir::Type ty, | ||||||
| 235 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 236 | if (box.getType().isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 237 | auto pty = mlir::LLVM::LLVMPointerType::get(ty); | ||||||
| 238 | mlir::LLVM::GEPOp p = genGEP(loc, pty, rewriter, box, 0, | ||||||
| 239 | static_cast<int>(kDimsPosInBox), dim, off); | ||||||
| 240 | auto loadOp = rewriter.create<mlir::LLVM::LoadOp>(loc, ty, p); | ||||||
| 241 | attachTBAATag(loadOp, boxTy, nullptr, p); | ||||||
| 242 | return loadOp; | ||||||
| 243 | } | ||||||
| 244 | return rewriter.create<mlir::LLVM::ExtractValueOp>( | ||||||
| 245 | loc, box, llvm::ArrayRef<std::int64_t>{kDimsPosInBox, dim, off}); | ||||||
| 246 | } | ||||||
| 247 | |||||||
| 248 | mlir::Value | ||||||
| 249 | getStrideFromBox(mlir::Location loc, mlir::Type boxTy, mlir::Value box, | ||||||
| 250 | unsigned dim, | ||||||
| 251 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 252 | auto idxTy = lowerTy().indexType(); | ||||||
| 253 | return getDimFieldFromBox(loc, boxTy, box, dim, kDimStridePos, idxTy, | ||||||
| 254 | rewriter); | ||||||
| 255 | } | ||||||
| 256 | |||||||
| 257 | /// Read base address from a fir.box. Returned address has type ty. | ||||||
| 258 | mlir::Value | ||||||
| 259 | getBaseAddrFromBox(mlir::Location loc, mlir::Type resultTy, mlir::Type boxTy, | ||||||
| 260 | mlir::Value box, | ||||||
| 261 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 262 | return getValueFromBox(loc, boxTy, box, resultTy, rewriter, kAddrPosInBox); | ||||||
| 263 | } | ||||||
| 264 | |||||||
| 265 | mlir::Value | ||||||
| 266 | getElementSizeFromBox(mlir::Location loc, mlir::Type resultTy, | ||||||
| 267 | mlir::Type boxTy, mlir::Value box, | ||||||
| 268 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 269 | return getValueFromBox(loc, boxTy, box, resultTy, rewriter, | ||||||
| 270 | kElemLenPosInBox); | ||||||
| 271 | } | ||||||
| 272 | |||||||
| 273 | // Get the element type given an LLVM type that is of the form | ||||||
| 274 | // [llvm.ptr](array|struct|vector)+ and the provided indexes. | ||||||
| 275 | static mlir::Type getBoxEleTy(mlir::Type type, | ||||||
| 276 | llvm::ArrayRef<std::int64_t> indexes) { | ||||||
| 277 | if (auto t = type.dyn_cast<mlir::LLVM::LLVMPointerType>()) | ||||||
| 278 | type = t.getElementType(); | ||||||
| 279 | for (unsigned i : indexes) { | ||||||
| 280 | if (auto t = type.dyn_cast<mlir::LLVM::LLVMStructType>()) { | ||||||
| 281 | assert(!t.isOpaque() && i < t.getBody().size())(static_cast <bool> (!t.isOpaque() && i < t. getBody().size()) ? void (0) : __assert_fail ("!t.isOpaque() && i < t.getBody().size()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 281, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 282 | type = t.getBody()[i]; | ||||||
| 283 | } else if (auto t = type.dyn_cast<mlir::LLVM::LLVMArrayType>()) { | ||||||
| 284 | type = t.getElementType(); | ||||||
| 285 | } else if (auto t = type.dyn_cast<mlir::VectorType>()) { | ||||||
| 286 | type = t.getElementType(); | ||||||
| 287 | } else { | ||||||
| 288 | fir::emitFatalError(mlir::UnknownLoc::get(type.getContext()), | ||||||
| 289 | "request for invalid box element type"); | ||||||
| 290 | } | ||||||
| 291 | } | ||||||
| 292 | return type; | ||||||
| 293 | } | ||||||
| 294 | |||||||
| 295 | // Return LLVM type of the base address given the LLVM type | ||||||
| 296 | // of the related descriptor (lowered fir.box type). | ||||||
| 297 | static mlir::Type getBaseAddrTypeFromBox(mlir::Type type) { | ||||||
| 298 | return getBoxEleTy(type, {kAddrPosInBox}); | ||||||
| 299 | } | ||||||
| 300 | |||||||
| 301 | /// Read the address of the type descriptor from a box. | ||||||
| 302 | mlir::Value | ||||||
| 303 | loadTypeDescAddress(mlir::Location loc, mlir::Type boxTy, mlir::Value box, | ||||||
| 304 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 305 | unsigned typeDescFieldId = getTypeDescFieldId(boxTy); | ||||||
| 306 | mlir::Type tdescType = lowerTy().convertTypeDescType(rewriter.getContext()); | ||||||
| 307 | return getValueFromBox(loc, boxTy, box, tdescType, rewriter, | ||||||
| 308 | typeDescFieldId); | ||||||
| 309 | } | ||||||
| 310 | |||||||
| 311 | // Load the attribute from the \p box and perform a check against \p maskValue | ||||||
| 312 | // The final comparison is implemented as `(attribute & maskValue) != 0`. | ||||||
| 313 | mlir::Value genBoxAttributeCheck(mlir::Location loc, mlir::Type boxTy, | ||||||
| 314 | mlir::Value box, | ||||||
| 315 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 316 | unsigned maskValue) const { | ||||||
| 317 | mlir::Type attrTy = rewriter.getI32Type(); | ||||||
| 318 | mlir::Value attribute = | ||||||
| 319 | getValueFromBox(loc, boxTy, box, attrTy, rewriter, kAttributePosInBox); | ||||||
| 320 | mlir::LLVM::ConstantOp attrMask = | ||||||
| 321 | genConstantOffset(loc, rewriter, maskValue); | ||||||
| 322 | auto maskRes = | ||||||
| 323 | rewriter.create<mlir::LLVM::AndOp>(loc, attrTy, attribute, attrMask); | ||||||
| 324 | mlir::LLVM::ConstantOp c0 = genConstantOffset(loc, rewriter, 0); | ||||||
| 325 | return rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 326 | loc, mlir::LLVM::ICmpPredicate::ne, maskRes, c0); | ||||||
| 327 | } | ||||||
| 328 | |||||||
| 329 | template <typename... ARGS> | ||||||
| 330 | mlir::LLVM::GEPOp genGEP(mlir::Location loc, mlir::Type ty, | ||||||
| 331 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 332 | mlir::Value base, ARGS... args) const { | ||||||
| 333 | llvm::SmallVector<mlir::LLVM::GEPArg> cv = {args...}; | ||||||
| 334 | return rewriter.create<mlir::LLVM::GEPOp>(loc, ty, base, cv); | ||||||
| 335 | } | ||||||
| 336 | |||||||
| 337 | // Find the LLVMFuncOp in whose entry block the alloca should be inserted. | ||||||
| 338 | // The order to find the LLVMFuncOp is as follows: | ||||||
| 339 | // 1. The parent operation of the current block if it is a LLVMFuncOp. | ||||||
| 340 | // 2. The first ancestor that is a LLVMFuncOp. | ||||||
| 341 | mlir::LLVM::LLVMFuncOp | ||||||
| 342 | getFuncForAllocaInsert(mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 343 | mlir::Operation *parentOp = rewriter.getInsertionBlock()->getParentOp(); | ||||||
| 344 | return mlir::isa<mlir::LLVM::LLVMFuncOp>(parentOp) | ||||||
| 345 | ? mlir::cast<mlir::LLVM::LLVMFuncOp>(parentOp) | ||||||
| 346 | : parentOp->getParentOfType<mlir::LLVM::LLVMFuncOp>(); | ||||||
| 347 | } | ||||||
| 348 | |||||||
| 349 | // Generate an alloca of size 1 and type \p toTy. | ||||||
| 350 | mlir::LLVM::AllocaOp | ||||||
| 351 | genAllocaWithType(mlir::Location loc, mlir::Type toTy, unsigned alignment, | ||||||
| 352 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 353 | auto thisPt = rewriter.saveInsertionPoint(); | ||||||
| 354 | mlir::LLVM::LLVMFuncOp func = getFuncForAllocaInsert(rewriter); | ||||||
| 355 | rewriter.setInsertionPointToStart(&func.front()); | ||||||
| 356 | auto size = genI32Constant(loc, rewriter, 1); | ||||||
| 357 | auto al = rewriter.create<mlir::LLVM::AllocaOp>(loc, toTy, size, alignment); | ||||||
| 358 | rewriter.restoreInsertionPoint(thisPt); | ||||||
| 359 | return al; | ||||||
| 360 | } | ||||||
| 361 | |||||||
| 362 | fir::LLVMTypeConverter &lowerTy() const { | ||||||
| 363 | return *static_cast<fir::LLVMTypeConverter *>(this->getTypeConverter()); | ||||||
| 364 | } | ||||||
| 365 | |||||||
| 366 | void attachTBAATag(mlir::LLVM::AliasAnalysisOpInterface op, | ||||||
| 367 | mlir::Type baseFIRType, mlir::Type accessFIRType, | ||||||
| 368 | mlir::LLVM::GEPOp gep) const { | ||||||
| 369 | lowerTy().attachTBAATag(op, baseFIRType, accessFIRType, gep); | ||||||
| 370 | } | ||||||
| 371 | |||||||
| 372 | const fir::FIRToLLVMPassOptions &options; | ||||||
| 373 | }; | ||||||
| 374 | |||||||
| 375 | /// FIR conversion pattern template | ||||||
| 376 | template <typename FromOp> | ||||||
| 377 | class FIROpAndTypeConversion : public FIROpConversion<FromOp> { | ||||||
| 378 | public: | ||||||
| 379 | using FIROpConversion<FromOp>::FIROpConversion; | ||||||
| 380 | using OpAdaptor = typename FromOp::Adaptor; | ||||||
| 381 | |||||||
| 382 | mlir::LogicalResult | ||||||
| 383 | matchAndRewrite(FromOp op, OpAdaptor adaptor, | ||||||
| 384 | mlir::ConversionPatternRewriter &rewriter) const final { | ||||||
| 385 | mlir::Type ty = this->convertType(op.getType()); | ||||||
| 386 | return doRewrite(op, ty, adaptor, rewriter); | ||||||
| 387 | } | ||||||
| 388 | |||||||
| 389 | virtual mlir::LogicalResult | ||||||
| 390 | doRewrite(FromOp addr, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 391 | mlir::ConversionPatternRewriter &rewriter) const = 0; | ||||||
| 392 | }; | ||||||
| 393 | } // namespace | ||||||
| 394 | |||||||
| 395 | namespace { | ||||||
| 396 | /// Lower `fir.address_of` operation to `llvm.address_of` operation. | ||||||
| 397 | struct AddrOfOpConversion : public FIROpConversion<fir::AddrOfOp> { | ||||||
| 398 | using FIROpConversion::FIROpConversion; | ||||||
| 399 | |||||||
| 400 | mlir::LogicalResult | ||||||
| 401 | matchAndRewrite(fir::AddrOfOp addr, OpAdaptor adaptor, | ||||||
| 402 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 403 | auto ty = convertType(addr.getType()); | ||||||
| 404 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>( | ||||||
| 405 | addr, ty, addr.getSymbol().getRootReference().getValue()); | ||||||
| 406 | return mlir::success(); | ||||||
| 407 | } | ||||||
| 408 | }; | ||||||
| 409 | } // namespace | ||||||
| 410 | |||||||
| 411 | /// Lookup the function to compute the memory size of this parametric derived | ||||||
| 412 | /// type. The size of the object may depend on the LEN type parameters of the | ||||||
| 413 | /// derived type. | ||||||
| 414 | static mlir::LLVM::LLVMFuncOp | ||||||
| 415 | getDependentTypeMemSizeFn(fir::RecordType recTy, fir::AllocaOp op, | ||||||
| 416 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 417 | auto module = op->getParentOfType<mlir::ModuleOp>(); | ||||||
| 418 | std::string name = recTy.getName().str() + "P.mem.size"; | ||||||
| 419 | if (auto memSizeFunc = module.lookupSymbol<mlir::LLVM::LLVMFuncOp>(name)) | ||||||
| 420 | return memSizeFunc; | ||||||
| 421 | TODO(op.getLoc(), "did not find allocation function")do { fir::emitFatalError(op.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "421" ": not yet implemented: ") + llvm::Twine("did not find allocation function" ), false); } while (false); | ||||||
| 422 | } | ||||||
| 423 | |||||||
| 424 | // Compute the alloc scale size (constant factors encoded in the array type). | ||||||
| 425 | // We do this for arrays without a constant interior or arrays of character with | ||||||
| 426 | // dynamic length arrays, since those are the only ones that get decayed to a | ||||||
| 427 | // pointer to the element type. | ||||||
| 428 | template <typename OP> | ||||||
| 429 | static mlir::Value | ||||||
| 430 | genAllocationScaleSize(OP op, mlir::Type ity, | ||||||
| 431 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 432 | mlir::Location loc = op.getLoc(); | ||||||
| 433 | mlir::Type dataTy = op.getInType(); | ||||||
| 434 | auto seqTy = dataTy.dyn_cast<fir::SequenceType>(); | ||||||
| 435 | fir::SequenceType::Extent constSize = 1; | ||||||
| 436 | if (seqTy) { | ||||||
| 437 | int constRows = seqTy.getConstantRows(); | ||||||
| 438 | const fir::SequenceType::ShapeRef &shape = seqTy.getShape(); | ||||||
| 439 | if (constRows != static_cast<int>(shape.size())) { | ||||||
| 440 | for (auto extent : shape) { | ||||||
| 441 | if (constRows-- > 0) | ||||||
| 442 | continue; | ||||||
| 443 | if (extent != fir::SequenceType::getUnknownExtent()) | ||||||
| 444 | constSize *= extent; | ||||||
| 445 | } | ||||||
| 446 | } | ||||||
| 447 | } | ||||||
| 448 | |||||||
| 449 | if (constSize != 1) { | ||||||
| 450 | mlir::Value constVal{ | ||||||
| 451 | genConstantIndex(loc, ity, rewriter, constSize).getResult()}; | ||||||
| 452 | return constVal; | ||||||
| 453 | } | ||||||
| 454 | return nullptr; | ||||||
| 455 | } | ||||||
| 456 | |||||||
| 457 | namespace { | ||||||
| 458 | /// convert to LLVM IR dialect `alloca` | ||||||
| 459 | struct AllocaOpConversion : public FIROpConversion<fir::AllocaOp> { | ||||||
| 460 | using FIROpConversion::FIROpConversion; | ||||||
| 461 | |||||||
| 462 | mlir::LogicalResult | ||||||
| 463 | matchAndRewrite(fir::AllocaOp alloc, OpAdaptor adaptor, | ||||||
| 464 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 465 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 466 | auto loc = alloc.getLoc(); | ||||||
| 467 | mlir::Type ity = lowerTy().indexType(); | ||||||
| 468 | unsigned i = 0; | ||||||
| 469 | mlir::Value size = genConstantIndex(loc, ity, rewriter, 1).getResult(); | ||||||
| 470 | mlir::Type ty = convertType(alloc.getType()); | ||||||
| 471 | mlir::Type resultTy = ty; | ||||||
| 472 | if (alloc.hasLenParams()) { | ||||||
| 473 | unsigned end = alloc.numLenParams(); | ||||||
| 474 | llvm::SmallVector<mlir::Value> lenParams; | ||||||
| 475 | for (; i < end; ++i) | ||||||
| 476 | lenParams.push_back(operands[i]); | ||||||
| 477 | mlir::Type scalarType = fir::unwrapSequenceType(alloc.getInType()); | ||||||
| 478 | if (auto chrTy = scalarType.dyn_cast<fir::CharacterType>()) { | ||||||
| 479 | fir::CharacterType rawCharTy = fir::CharacterType::getUnknownLen( | ||||||
| 480 | chrTy.getContext(), chrTy.getFKind()); | ||||||
| 481 | ty = mlir::LLVM::LLVMPointerType::get(convertType(rawCharTy)); | ||||||
| 482 | assert(end == 1)(static_cast <bool> (end == 1) ? void (0) : __assert_fail ("end == 1", "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 482, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 483 | size = integerCast(loc, rewriter, ity, lenParams[0]); | ||||||
| 484 | } else if (auto recTy = scalarType.dyn_cast<fir::RecordType>()) { | ||||||
| 485 | mlir::LLVM::LLVMFuncOp memSizeFn = | ||||||
| 486 | getDependentTypeMemSizeFn(recTy, alloc, rewriter); | ||||||
| 487 | if (!memSizeFn) | ||||||
| 488 | emitError(loc, "did not find allocation function"); | ||||||
| 489 | mlir::NamedAttribute attr = rewriter.getNamedAttr( | ||||||
| 490 | "callee", mlir::SymbolRefAttr::get(memSizeFn)); | ||||||
| 491 | auto call = rewriter.create<mlir::LLVM::CallOp>( | ||||||
| 492 | loc, ity, lenParams, llvm::ArrayRef<mlir::NamedAttribute>{attr}); | ||||||
| 493 | size = call.getResult(); | ||||||
| 494 | ty = ::getVoidPtrType(alloc.getContext()); | ||||||
| 495 | } else { | ||||||
| 496 | return emitError(loc, "unexpected type ") | ||||||
| 497 | << scalarType << " with type parameters"; | ||||||
| 498 | } | ||||||
| 499 | } | ||||||
| 500 | if (auto scaleSize = genAllocationScaleSize(alloc, ity, rewriter)) | ||||||
| 501 | size = rewriter.create<mlir::LLVM::MulOp>(loc, ity, size, scaleSize); | ||||||
| 502 | if (alloc.hasShapeOperands()) { | ||||||
| 503 | unsigned end = operands.size(); | ||||||
| 504 | for (; i < end; ++i) | ||||||
| 505 | size = rewriter.create<mlir::LLVM::MulOp>( | ||||||
| 506 | loc, ity, size, integerCast(loc, rewriter, ity, operands[i])); | ||||||
| 507 | } | ||||||
| 508 | if (ty == resultTy) { | ||||||
| 509 | // Do not emit the bitcast if ty and resultTy are the same. | ||||||
| 510 | rewriter.replaceOpWithNewOp<mlir::LLVM::AllocaOp>(alloc, ty, size, | ||||||
| 511 | alloc->getAttrs()); | ||||||
| 512 | } else { | ||||||
| 513 | auto al = rewriter.create<mlir::LLVM::AllocaOp>(loc, ty, size, | ||||||
| 514 | alloc->getAttrs()); | ||||||
| 515 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(alloc, resultTy, al); | ||||||
| 516 | } | ||||||
| 517 | return mlir::success(); | ||||||
| 518 | } | ||||||
| 519 | }; | ||||||
| 520 | } // namespace | ||||||
| 521 | |||||||
| 522 | namespace { | ||||||
| 523 | /// Lower `fir.box_addr` to the sequence of operations to extract the first | ||||||
| 524 | /// element of the box. | ||||||
| 525 | struct BoxAddrOpConversion : public FIROpConversion<fir::BoxAddrOp> { | ||||||
| 526 | using FIROpConversion::FIROpConversion; | ||||||
| 527 | |||||||
| 528 | mlir::LogicalResult | ||||||
| 529 | matchAndRewrite(fir::BoxAddrOp boxaddr, OpAdaptor adaptor, | ||||||
| 530 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 531 | mlir::Value a = adaptor.getOperands()[0]; | ||||||
| 532 | auto loc = boxaddr.getLoc(); | ||||||
| 533 | mlir::Type ty = convertType(boxaddr.getType()); | ||||||
| 534 | if (auto argty = boxaddr.getVal().getType().dyn_cast<fir::BaseBoxType>()) { | ||||||
| 535 | rewriter.replaceOp(boxaddr, | ||||||
| 536 | getBaseAddrFromBox(loc, ty, argty, a, rewriter)); | ||||||
| 537 | } else { | ||||||
| 538 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractValueOp>(boxaddr, a, 0); | ||||||
| 539 | } | ||||||
| 540 | return mlir::success(); | ||||||
| 541 | } | ||||||
| 542 | }; | ||||||
| 543 | |||||||
| 544 | /// Convert `!fir.boxchar_len` to `!llvm.extractvalue` for the 2nd part of the | ||||||
| 545 | /// boxchar. | ||||||
| 546 | struct BoxCharLenOpConversion : public FIROpConversion<fir::BoxCharLenOp> { | ||||||
| 547 | using FIROpConversion::FIROpConversion; | ||||||
| 548 | |||||||
| 549 | mlir::LogicalResult | ||||||
| 550 | matchAndRewrite(fir::BoxCharLenOp boxCharLen, OpAdaptor adaptor, | ||||||
| 551 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 552 | mlir::Value boxChar = adaptor.getOperands()[0]; | ||||||
| 553 | mlir::Location loc = boxChar.getLoc(); | ||||||
| 554 | mlir::Type returnValTy = boxCharLen.getResult().getType(); | ||||||
| 555 | |||||||
| 556 | constexpr int boxcharLenIdx = 1; | ||||||
| 557 | auto len = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, boxChar, | ||||||
| 558 | boxcharLenIdx); | ||||||
| 559 | mlir::Value lenAfterCast = integerCast(loc, rewriter, returnValTy, len); | ||||||
| 560 | rewriter.replaceOp(boxCharLen, lenAfterCast); | ||||||
| 561 | |||||||
| 562 | return mlir::success(); | ||||||
| 563 | } | ||||||
| 564 | }; | ||||||
| 565 | |||||||
| 566 | /// Lower `fir.box_dims` to a sequence of operations to extract the requested | ||||||
| 567 | /// dimension information from the boxed value. | ||||||
| 568 | /// Result in a triple set of GEPs and loads. | ||||||
| 569 | struct BoxDimsOpConversion : public FIROpConversion<fir::BoxDimsOp> { | ||||||
| 570 | using FIROpConversion::FIROpConversion; | ||||||
| 571 | |||||||
| 572 | mlir::LogicalResult | ||||||
| 573 | matchAndRewrite(fir::BoxDimsOp boxdims, OpAdaptor adaptor, | ||||||
| 574 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 575 | llvm::SmallVector<mlir::Type, 3> resultTypes = { | ||||||
| 576 | convertType(boxdims.getResult(0).getType()), | ||||||
| 577 | convertType(boxdims.getResult(1).getType()), | ||||||
| 578 | convertType(boxdims.getResult(2).getType()), | ||||||
| 579 | }; | ||||||
| 580 | auto results = getDimsFromBox( | ||||||
| 581 | boxdims.getLoc(), resultTypes, boxdims.getVal().getType(), | ||||||
| 582 | adaptor.getOperands()[0], adaptor.getOperands()[1], rewriter); | ||||||
| 583 | rewriter.replaceOp(boxdims, results); | ||||||
| 584 | return mlir::success(); | ||||||
| 585 | } | ||||||
| 586 | }; | ||||||
| 587 | |||||||
| 588 | /// Lower `fir.box_elesize` to a sequence of operations ro extract the size of | ||||||
| 589 | /// an element in the boxed value. | ||||||
| 590 | struct BoxEleSizeOpConversion : public FIROpConversion<fir::BoxEleSizeOp> { | ||||||
| 591 | using FIROpConversion::FIROpConversion; | ||||||
| 592 | |||||||
| 593 | mlir::LogicalResult | ||||||
| 594 | matchAndRewrite(fir::BoxEleSizeOp boxelesz, OpAdaptor adaptor, | ||||||
| 595 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 596 | mlir::Value box = adaptor.getOperands()[0]; | ||||||
| 597 | auto loc = boxelesz.getLoc(); | ||||||
| 598 | auto ty = convertType(boxelesz.getType()); | ||||||
| 599 | auto elemSize = getElementSizeFromBox(loc, ty, boxelesz.getVal().getType(), | ||||||
| 600 | box, rewriter); | ||||||
| 601 | rewriter.replaceOp(boxelesz, elemSize); | ||||||
| 602 | return mlir::success(); | ||||||
| 603 | } | ||||||
| 604 | }; | ||||||
| 605 | |||||||
| 606 | /// Lower `fir.box_isalloc` to a sequence of operations to determine if the | ||||||
| 607 | /// boxed value was from an ALLOCATABLE entity. | ||||||
| 608 | struct BoxIsAllocOpConversion : public FIROpConversion<fir::BoxIsAllocOp> { | ||||||
| 609 | using FIROpConversion::FIROpConversion; | ||||||
| 610 | |||||||
| 611 | mlir::LogicalResult | ||||||
| 612 | matchAndRewrite(fir::BoxIsAllocOp boxisalloc, OpAdaptor adaptor, | ||||||
| 613 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 614 | mlir::Value box = adaptor.getOperands()[0]; | ||||||
| 615 | auto loc = boxisalloc.getLoc(); | ||||||
| 616 | mlir::Value check = genBoxAttributeCheck(loc, boxisalloc.getVal().getType(), | ||||||
| 617 | box, rewriter, kAttrAllocatable); | ||||||
| 618 | rewriter.replaceOp(boxisalloc, check); | ||||||
| 619 | return mlir::success(); | ||||||
| 620 | } | ||||||
| 621 | }; | ||||||
| 622 | |||||||
| 623 | /// Lower `fir.box_isarray` to a sequence of operations to determine if the | ||||||
| 624 | /// boxed is an array. | ||||||
| 625 | struct BoxIsArrayOpConversion : public FIROpConversion<fir::BoxIsArrayOp> { | ||||||
| 626 | using FIROpConversion::FIROpConversion; | ||||||
| 627 | |||||||
| 628 | mlir::LogicalResult | ||||||
| 629 | matchAndRewrite(fir::BoxIsArrayOp boxisarray, OpAdaptor adaptor, | ||||||
| 630 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 631 | mlir::Value a = adaptor.getOperands()[0]; | ||||||
| 632 | auto loc = boxisarray.getLoc(); | ||||||
| 633 | auto rank = getValueFromBox(loc, boxisarray.getVal().getType(), a, | ||||||
| 634 | rewriter.getI32Type(), rewriter, kRankPosInBox); | ||||||
| 635 | auto c0 = genConstantOffset(loc, rewriter, 0); | ||||||
| 636 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( | ||||||
| 637 | boxisarray, mlir::LLVM::ICmpPredicate::ne, rank, c0); | ||||||
| 638 | return mlir::success(); | ||||||
| 639 | } | ||||||
| 640 | }; | ||||||
| 641 | |||||||
| 642 | /// Lower `fir.box_isptr` to a sequence of operations to determined if the | ||||||
| 643 | /// boxed value was from a POINTER entity. | ||||||
| 644 | struct BoxIsPtrOpConversion : public FIROpConversion<fir::BoxIsPtrOp> { | ||||||
| 645 | using FIROpConversion::FIROpConversion; | ||||||
| 646 | |||||||
| 647 | mlir::LogicalResult | ||||||
| 648 | matchAndRewrite(fir::BoxIsPtrOp boxisptr, OpAdaptor adaptor, | ||||||
| 649 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 650 | mlir::Value box = adaptor.getOperands()[0]; | ||||||
| 651 | auto loc = boxisptr.getLoc(); | ||||||
| 652 | mlir::Value check = genBoxAttributeCheck(loc, boxisptr.getVal().getType(), | ||||||
| 653 | box, rewriter, kAttrPointer); | ||||||
| 654 | rewriter.replaceOp(boxisptr, check); | ||||||
| 655 | return mlir::success(); | ||||||
| 656 | } | ||||||
| 657 | }; | ||||||
| 658 | |||||||
| 659 | /// Lower `fir.box_rank` to the sequence of operation to extract the rank from | ||||||
| 660 | /// the box. | ||||||
| 661 | struct BoxRankOpConversion : public FIROpConversion<fir::BoxRankOp> { | ||||||
| 662 | using FIROpConversion::FIROpConversion; | ||||||
| 663 | |||||||
| 664 | mlir::LogicalResult | ||||||
| 665 | matchAndRewrite(fir::BoxRankOp boxrank, OpAdaptor adaptor, | ||||||
| 666 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 667 | mlir::Value a = adaptor.getOperands()[0]; | ||||||
| 668 | auto loc = boxrank.getLoc(); | ||||||
| 669 | mlir::Type ty = convertType(boxrank.getType()); | ||||||
| 670 | auto result = getValueFromBox(loc, boxrank.getVal().getType(), a, ty, | ||||||
| 671 | rewriter, kRankPosInBox); | ||||||
| 672 | rewriter.replaceOp(boxrank, result); | ||||||
| 673 | return mlir::success(); | ||||||
| 674 | } | ||||||
| 675 | }; | ||||||
| 676 | |||||||
| 677 | /// Lower `fir.boxproc_host` operation. Extracts the host pointer from the | ||||||
| 678 | /// boxproc. | ||||||
| 679 | /// TODO: Part of supporting Fortran 2003 procedure pointers. | ||||||
| 680 | struct BoxProcHostOpConversion : public FIROpConversion<fir::BoxProcHostOp> { | ||||||
| 681 | using FIROpConversion::FIROpConversion; | ||||||
| 682 | |||||||
| 683 | mlir::LogicalResult | ||||||
| 684 | matchAndRewrite(fir::BoxProcHostOp boxprochost, OpAdaptor adaptor, | ||||||
| 685 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 686 | TODO(boxprochost.getLoc(), "fir.boxproc_host codegen")do { fir::emitFatalError(boxprochost.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "686" ": not yet implemented: ") + llvm::Twine("fir.boxproc_host codegen" ), false); } while (false); | ||||||
| 687 | return mlir::failure(); | ||||||
| 688 | } | ||||||
| 689 | }; | ||||||
| 690 | |||||||
| 691 | /// Lower `fir.box_tdesc` to the sequence of operations to extract the type | ||||||
| 692 | /// descriptor from the box. | ||||||
| 693 | struct BoxTypeDescOpConversion : public FIROpConversion<fir::BoxTypeDescOp> { | ||||||
| 694 | using FIROpConversion::FIROpConversion; | ||||||
| 695 | |||||||
| 696 | mlir::LogicalResult | ||||||
| 697 | matchAndRewrite(fir::BoxTypeDescOp boxtypedesc, OpAdaptor adaptor, | ||||||
| 698 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 699 | mlir::Value box = adaptor.getOperands()[0]; | ||||||
| 700 | auto typeDescAddr = loadTypeDescAddress( | ||||||
| 701 | boxtypedesc.getLoc(), boxtypedesc.getBox().getType(), box, rewriter); | ||||||
| 702 | rewriter.replaceOp(boxtypedesc, typeDescAddr); | ||||||
| 703 | return mlir::success(); | ||||||
| 704 | } | ||||||
| 705 | }; | ||||||
| 706 | |||||||
| 707 | /// Lower `fir.box_typecode` to a sequence of operations to extract the type | ||||||
| 708 | /// code in the boxed value. | ||||||
| 709 | struct BoxTypeCodeOpConversion : public FIROpConversion<fir::BoxTypeCodeOp> { | ||||||
| 710 | using FIROpConversion::FIROpConversion; | ||||||
| 711 | |||||||
| 712 | mlir::LogicalResult | ||||||
| 713 | matchAndRewrite(fir::BoxTypeCodeOp op, OpAdaptor adaptor, | ||||||
| 714 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 715 | mlir::Value box = adaptor.getOperands()[0]; | ||||||
| 716 | auto loc = box.getLoc(); | ||||||
| 717 | auto ty = convertType(op.getType()); | ||||||
| 718 | auto typeCode = getValueFromBox(loc, op.getBox().getType(), box, ty, | ||||||
| 719 | rewriter, kTypePosInBox); | ||||||
| 720 | rewriter.replaceOp(op, typeCode); | ||||||
| 721 | return mlir::success(); | ||||||
| 722 | } | ||||||
| 723 | }; | ||||||
| 724 | |||||||
| 725 | /// Lower `fir.string_lit` to LLVM IR dialect operation. | ||||||
| 726 | struct StringLitOpConversion : public FIROpConversion<fir::StringLitOp> { | ||||||
| 727 | using FIROpConversion::FIROpConversion; | ||||||
| 728 | |||||||
| 729 | mlir::LogicalResult | ||||||
| 730 | matchAndRewrite(fir::StringLitOp constop, OpAdaptor adaptor, | ||||||
| 731 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 732 | auto ty = convertType(constop.getType()); | ||||||
| 733 | auto attr = constop.getValue(); | ||||||
| 734 | if (attr.isa<mlir::StringAttr>()) { | ||||||
| 735 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>(constop, ty, attr); | ||||||
| 736 | return mlir::success(); | ||||||
| 737 | } | ||||||
| 738 | |||||||
| 739 | auto charTy = constop.getType().cast<fir::CharacterType>(); | ||||||
| 740 | unsigned bits = lowerTy().characterBitsize(charTy); | ||||||
| 741 | mlir::Type intTy = rewriter.getIntegerType(bits); | ||||||
| 742 | mlir::Location loc = constop.getLoc(); | ||||||
| 743 | mlir::Value cst = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 744 | if (auto arr = attr.dyn_cast<mlir::DenseElementsAttr>()) { | ||||||
| 745 | cst = rewriter.create<mlir::LLVM::ConstantOp>(loc, ty, arr); | ||||||
| 746 | } else if (auto arr = attr.dyn_cast<mlir::ArrayAttr>()) { | ||||||
| 747 | for (auto a : llvm::enumerate(arr.getValue())) { | ||||||
| 748 | // convert each character to a precise bitsize | ||||||
| 749 | auto elemAttr = mlir::IntegerAttr::get( | ||||||
| 750 | intTy, | ||||||
| 751 | a.value().cast<mlir::IntegerAttr>().getValue().zextOrTrunc(bits)); | ||||||
| 752 | auto elemCst = | ||||||
| 753 | rewriter.create<mlir::LLVM::ConstantOp>(loc, intTy, elemAttr); | ||||||
| 754 | cst = rewriter.create<mlir::LLVM::InsertValueOp>(loc, cst, elemCst, | ||||||
| 755 | a.index()); | ||||||
| 756 | } | ||||||
| 757 | } else { | ||||||
| 758 | return mlir::failure(); | ||||||
| 759 | } | ||||||
| 760 | rewriter.replaceOp(constop, cst); | ||||||
| 761 | return mlir::success(); | ||||||
| 762 | } | ||||||
| 763 | }; | ||||||
| 764 | |||||||
| 765 | /// `fir.call` -> `llvm.call` | ||||||
| 766 | struct CallOpConversion : public FIROpConversion<fir::CallOp> { | ||||||
| 767 | using FIROpConversion::FIROpConversion; | ||||||
| 768 | |||||||
| 769 | mlir::LogicalResult | ||||||
| 770 | matchAndRewrite(fir::CallOp call, OpAdaptor adaptor, | ||||||
| 771 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 772 | llvm::SmallVector<mlir::Type> resultTys; | ||||||
| 773 | for (auto r : call.getResults()) | ||||||
| 774 | resultTys.push_back(convertType(r.getType())); | ||||||
| 775 | // Convert arith::FastMathFlagsAttr to LLVM::FastMathFlagsAttr. | ||||||
| 776 | mlir::arith::AttrConvertFastMathToLLVM<fir::CallOp, mlir::LLVM::CallOp> | ||||||
| 777 | attrConvert(call); | ||||||
| 778 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( | ||||||
| 779 | call, resultTys, adaptor.getOperands(), attrConvert.getAttrs()); | ||||||
| 780 | return mlir::success(); | ||||||
| 781 | } | ||||||
| 782 | }; | ||||||
| 783 | } // namespace | ||||||
| 784 | |||||||
| 785 | static mlir::Type getComplexEleTy(mlir::Type complex) { | ||||||
| 786 | if (auto cc = complex.dyn_cast<mlir::ComplexType>()) | ||||||
| 787 | return cc.getElementType(); | ||||||
| 788 | return complex.cast<fir::ComplexType>().getElementType(); | ||||||
| 789 | } | ||||||
| 790 | |||||||
| 791 | namespace { | ||||||
| 792 | /// Compare complex values | ||||||
| 793 | /// | ||||||
| 794 | /// Per 10.1, the only comparisons available are .EQ. (oeq) and .NE. (une). | ||||||
| 795 | /// | ||||||
| 796 | /// For completeness, all other comparison are done on the real component only. | ||||||
| 797 | struct CmpcOpConversion : public FIROpConversion<fir::CmpcOp> { | ||||||
| 798 | using FIROpConversion::FIROpConversion; | ||||||
| 799 | |||||||
| 800 | mlir::LogicalResult | ||||||
| 801 | matchAndRewrite(fir::CmpcOp cmp, OpAdaptor adaptor, | ||||||
| 802 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 803 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 804 | mlir::Type resTy = convertType(cmp.getType()); | ||||||
| 805 | mlir::Location loc = cmp.getLoc(); | ||||||
| 806 | llvm::SmallVector<mlir::Value, 2> rp = { | ||||||
| 807 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[0], 0), | ||||||
| 808 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[1], 0)}; | ||||||
| 809 | auto rcp = | ||||||
| 810 | rewriter.create<mlir::LLVM::FCmpOp>(loc, resTy, rp, cmp->getAttrs()); | ||||||
| 811 | llvm::SmallVector<mlir::Value, 2> ip = { | ||||||
| 812 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[0], 1), | ||||||
| 813 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, operands[1], 1)}; | ||||||
| 814 | auto icp = | ||||||
| 815 | rewriter.create<mlir::LLVM::FCmpOp>(loc, resTy, ip, cmp->getAttrs()); | ||||||
| 816 | llvm::SmallVector<mlir::Value, 2> cp = {rcp, icp}; | ||||||
| 817 | switch (cmp.getPredicate()) { | ||||||
| 818 | case mlir::arith::CmpFPredicate::OEQ: // .EQ. | ||||||
| 819 | rewriter.replaceOpWithNewOp<mlir::LLVM::AndOp>(cmp, resTy, cp); | ||||||
| 820 | break; | ||||||
| 821 | case mlir::arith::CmpFPredicate::UNE: // .NE. | ||||||
| 822 | rewriter.replaceOpWithNewOp<mlir::LLVM::OrOp>(cmp, resTy, cp); | ||||||
| 823 | break; | ||||||
| 824 | default: | ||||||
| 825 | rewriter.replaceOp(cmp, rcp.getResult()); | ||||||
| 826 | break; | ||||||
| 827 | } | ||||||
| 828 | return mlir::success(); | ||||||
| 829 | } | ||||||
| 830 | }; | ||||||
| 831 | |||||||
| 832 | /// Lower complex constants | ||||||
| 833 | struct ConstcOpConversion : public FIROpConversion<fir::ConstcOp> { | ||||||
| 834 | using FIROpConversion::FIROpConversion; | ||||||
| 835 | |||||||
| 836 | mlir::LogicalResult | ||||||
| 837 | matchAndRewrite(fir::ConstcOp conc, OpAdaptor, | ||||||
| 838 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 839 | mlir::Location loc = conc.getLoc(); | ||||||
| 840 | mlir::Type ty = convertType(conc.getType()); | ||||||
| 841 | mlir::Type ety = convertType(getComplexEleTy(conc.getType())); | ||||||
| 842 | auto realPart = rewriter.create<mlir::LLVM::ConstantOp>( | ||||||
| 843 | loc, ety, getValue(conc.getReal())); | ||||||
| 844 | auto imPart = rewriter.create<mlir::LLVM::ConstantOp>( | ||||||
| 845 | loc, ety, getValue(conc.getImaginary())); | ||||||
| 846 | auto undef = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 847 | auto setReal = | ||||||
| 848 | rewriter.create<mlir::LLVM::InsertValueOp>(loc, undef, realPart, 0); | ||||||
| 849 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(conc, setReal, | ||||||
| 850 | imPart, 1); | ||||||
| 851 | return mlir::success(); | ||||||
| 852 | } | ||||||
| 853 | |||||||
| 854 | inline llvm::APFloat getValue(mlir::Attribute attr) const { | ||||||
| 855 | return attr.cast<fir::RealAttr>().getValue(); | ||||||
| 856 | } | ||||||
| 857 | }; | ||||||
| 858 | |||||||
| 859 | /// convert value of from-type to value of to-type | ||||||
| 860 | struct ConvertOpConversion : public FIROpConversion<fir::ConvertOp> { | ||||||
| 861 | using FIROpConversion::FIROpConversion; | ||||||
| 862 | |||||||
| 863 | static bool isFloatingPointTy(mlir::Type ty) { | ||||||
| 864 | return ty.isa<mlir::FloatType>(); | ||||||
| 865 | } | ||||||
| 866 | |||||||
| 867 | mlir::LogicalResult | ||||||
| 868 | matchAndRewrite(fir::ConvertOp convert, OpAdaptor adaptor, | ||||||
| 869 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 870 | auto fromFirTy = convert.getValue().getType(); | ||||||
| 871 | auto toFirTy = convert.getRes().getType(); | ||||||
| 872 | auto fromTy = convertType(fromFirTy); | ||||||
| 873 | auto toTy = convertType(toFirTy); | ||||||
| 874 | mlir::Value op0 = adaptor.getOperands()[0]; | ||||||
| 875 | |||||||
| 876 | if (fromFirTy == toFirTy) { | ||||||
| 877 | rewriter.replaceOp(convert, op0); | ||||||
| 878 | return mlir::success(); | ||||||
| 879 | } | ||||||
| 880 | |||||||
| 881 | auto loc = convert.getLoc(); | ||||||
| 882 | auto i1Type = mlir::IntegerType::get(convert.getContext(), 1); | ||||||
| 883 | |||||||
| 884 | if (fromFirTy.isa<fir::LogicalType>() || toFirTy.isa<fir::LogicalType>()) { | ||||||
| 885 | // By specification fir::LogicalType value may be any number, | ||||||
| 886 | // where non-zero value represents .true. and zero value represents | ||||||
| 887 | // .false. | ||||||
| 888 | // | ||||||
| 889 | // integer<->logical conversion requires value normalization. | ||||||
| 890 | // Conversion from wide logical to narrow logical must set the result | ||||||
| 891 | // to non-zero iff the input is non-zero - the easiest way to implement | ||||||
| 892 | // it is to compare the input agains zero and set the result to | ||||||
| 893 | // the canonical 0/1. | ||||||
| 894 | // Conversion from narrow logical to wide logical may be implemented | ||||||
| 895 | // as a zero or sign extension of the input, but it may use value | ||||||
| 896 | // normalization as well. | ||||||
| 897 | if (!fromTy.isa<mlir::IntegerType>() || !toTy.isa<mlir::IntegerType>()) | ||||||
| 898 | return mlir::emitError(loc) | ||||||
| 899 | << "unsupported types for logical conversion: " << fromTy | ||||||
| 900 | << " -> " << toTy; | ||||||
| 901 | |||||||
| 902 | // Do folding for constant inputs. | ||||||
| 903 | if (auto constVal = getIfConstantIntValue(op0)) { | ||||||
| 904 | mlir::Value normVal = | ||||||
| 905 | genConstantIndex(loc, toTy, rewriter, *constVal ? 1 : 0); | ||||||
| 906 | rewriter.replaceOp(convert, normVal); | ||||||
| 907 | return mlir::success(); | ||||||
| 908 | } | ||||||
| 909 | |||||||
| 910 | // If the input is i1, then we can just zero extend it, and | ||||||
| 911 | // the result will be normalized. | ||||||
| 912 | if (fromTy == i1Type) { | ||||||
| 913 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, op0); | ||||||
| 914 | return mlir::success(); | ||||||
| 915 | } | ||||||
| 916 | |||||||
| 917 | // Compare the input with zero. | ||||||
| 918 | mlir::Value zero = genConstantIndex(loc, fromTy, rewriter, 0); | ||||||
| 919 | auto isTrue = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 920 | loc, mlir::LLVM::ICmpPredicate::ne, op0, zero); | ||||||
| 921 | |||||||
| 922 | // Zero extend the i1 isTrue result to the required type (unless it is i1 | ||||||
| 923 | // itself). | ||||||
| 924 | if (toTy != i1Type) | ||||||
| 925 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, isTrue); | ||||||
| 926 | else | ||||||
| 927 | rewriter.replaceOp(convert, isTrue.getResult()); | ||||||
| 928 | |||||||
| 929 | return mlir::success(); | ||||||
| 930 | } | ||||||
| 931 | |||||||
| 932 | if (fromTy == toTy) { | ||||||
| 933 | rewriter.replaceOp(convert, op0); | ||||||
| 934 | return mlir::success(); | ||||||
| 935 | } | ||||||
| 936 | auto convertFpToFp = [&](mlir::Value val, unsigned fromBits, | ||||||
| 937 | unsigned toBits, mlir::Type toTy) -> mlir::Value { | ||||||
| 938 | if (fromBits == toBits) { | ||||||
| 939 | // TODO: Converting between two floating-point representations with the | ||||||
| 940 | // same bitwidth is not allowed for now. | ||||||
| 941 | mlir::emitError(loc, | ||||||
| 942 | "cannot implicitly convert between two floating-point " | ||||||
| 943 | "representations of the same bitwidth"); | ||||||
| 944 | return {}; | ||||||
| 945 | } | ||||||
| 946 | if (fromBits > toBits) | ||||||
| 947 | return rewriter.create<mlir::LLVM::FPTruncOp>(loc, toTy, val); | ||||||
| 948 | return rewriter.create<mlir::LLVM::FPExtOp>(loc, toTy, val); | ||||||
| 949 | }; | ||||||
| 950 | // Complex to complex conversion. | ||||||
| 951 | if (fir::isa_complex(fromFirTy) && fir::isa_complex(toFirTy)) { | ||||||
| 952 | // Special case: handle the conversion of a complex such that both the | ||||||
| 953 | // real and imaginary parts are converted together. | ||||||
| 954 | auto ty = convertType(getComplexEleTy(convert.getValue().getType())); | ||||||
| 955 | auto rp = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, op0, 0); | ||||||
| 956 | auto ip = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, op0, 1); | ||||||
| 957 | auto nt = convertType(getComplexEleTy(convert.getRes().getType())); | ||||||
| 958 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(ty); | ||||||
| 959 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(nt); | ||||||
| 960 | auto rc = convertFpToFp(rp, fromBits, toBits, nt); | ||||||
| 961 | auto ic = convertFpToFp(ip, fromBits, toBits, nt); | ||||||
| 962 | auto un = rewriter.create<mlir::LLVM::UndefOp>(loc, toTy); | ||||||
| 963 | auto i1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, un, rc, 0); | ||||||
| 964 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(convert, i1, ic, | ||||||
| 965 | 1); | ||||||
| 966 | return mlir::success(); | ||||||
| 967 | } | ||||||
| 968 | |||||||
| 969 | // Floating point to floating point conversion. | ||||||
| 970 | if (isFloatingPointTy(fromTy)) { | ||||||
| 971 | if (isFloatingPointTy(toTy)) { | ||||||
| 972 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(fromTy); | ||||||
| 973 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(toTy); | ||||||
| 974 | auto v = convertFpToFp(op0, fromBits, toBits, toTy); | ||||||
| 975 | rewriter.replaceOp(convert, v); | ||||||
| 976 | return mlir::success(); | ||||||
| 977 | } | ||||||
| 978 | if (toTy.isa<mlir::IntegerType>()) { | ||||||
| 979 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPToSIOp>(convert, toTy, op0); | ||||||
| 980 | return mlir::success(); | ||||||
| 981 | } | ||||||
| 982 | } else if (fromTy.isa<mlir::IntegerType>()) { | ||||||
| 983 | // Integer to integer conversion. | ||||||
| 984 | if (toTy.isa<mlir::IntegerType>()) { | ||||||
| 985 | auto fromBits = mlir::LLVM::getPrimitiveTypeSizeInBits(fromTy); | ||||||
| 986 | auto toBits = mlir::LLVM::getPrimitiveTypeSizeInBits(toTy); | ||||||
| 987 | assert(fromBits != toBits)(static_cast <bool> (fromBits != toBits) ? void (0) : __assert_fail ("fromBits != toBits", "flang/lib/Optimizer/CodeGen/CodeGen.cpp" , 987, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 988 | if (fromBits > toBits) { | ||||||
| 989 | rewriter.replaceOpWithNewOp<mlir::LLVM::TruncOp>(convert, toTy, op0); | ||||||
| 990 | return mlir::success(); | ||||||
| 991 | } | ||||||
| 992 | if (fromFirTy == i1Type) { | ||||||
| 993 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(convert, toTy, op0); | ||||||
| 994 | return mlir::success(); | ||||||
| 995 | } | ||||||
| 996 | rewriter.replaceOpWithNewOp<mlir::LLVM::SExtOp>(convert, toTy, op0); | ||||||
| 997 | return mlir::success(); | ||||||
| 998 | } | ||||||
| 999 | // Integer to floating point conversion. | ||||||
| 1000 | if (isFloatingPointTy(toTy)) { | ||||||
| 1001 | rewriter.replaceOpWithNewOp<mlir::LLVM::SIToFPOp>(convert, toTy, op0); | ||||||
| 1002 | return mlir::success(); | ||||||
| 1003 | } | ||||||
| 1004 | // Integer to pointer conversion. | ||||||
| 1005 | if (toTy.isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 1006 | rewriter.replaceOpWithNewOp<mlir::LLVM::IntToPtrOp>(convert, toTy, op0); | ||||||
| 1007 | return mlir::success(); | ||||||
| 1008 | } | ||||||
| 1009 | } else if (fromTy.isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 1010 | // Pointer to integer conversion. | ||||||
| 1011 | if (toTy.isa<mlir::IntegerType>()) { | ||||||
| 1012 | rewriter.replaceOpWithNewOp<mlir::LLVM::PtrToIntOp>(convert, toTy, op0); | ||||||
| 1013 | return mlir::success(); | ||||||
| 1014 | } | ||||||
| 1015 | // Pointer to pointer conversion. | ||||||
| 1016 | if (toTy.isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 1017 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(convert, toTy, op0); | ||||||
| 1018 | return mlir::success(); | ||||||
| 1019 | } | ||||||
| 1020 | } | ||||||
| 1021 | return emitError(loc) << "cannot convert " << fromTy << " to " << toTy; | ||||||
| 1022 | } | ||||||
| 1023 | }; | ||||||
| 1024 | |||||||
| 1025 | /// `fir.disptach_table` operation has no specific CodeGen. The operation is | ||||||
| 1026 | /// only used to carry information during FIR to FIR passes. | ||||||
| 1027 | struct DispatchTableOpConversion | ||||||
| 1028 | : public FIROpConversion<fir::DispatchTableOp> { | ||||||
| 1029 | using FIROpConversion::FIROpConversion; | ||||||
| 1030 | |||||||
| 1031 | mlir::LogicalResult | ||||||
| 1032 | matchAndRewrite(fir::DispatchTableOp op, OpAdaptor, | ||||||
| 1033 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1034 | rewriter.eraseOp(op); | ||||||
| 1035 | return mlir::success(); | ||||||
| 1036 | } | ||||||
| 1037 | }; | ||||||
| 1038 | |||||||
| 1039 | /// `fir.dt_entry` operation has no specific CodeGen. The operation is only used | ||||||
| 1040 | /// to carry information during FIR to FIR passes. | ||||||
| 1041 | struct DTEntryOpConversion : public FIROpConversion<fir::DTEntryOp> { | ||||||
| 1042 | using FIROpConversion::FIROpConversion; | ||||||
| 1043 | |||||||
| 1044 | mlir::LogicalResult | ||||||
| 1045 | matchAndRewrite(fir::DTEntryOp op, OpAdaptor, | ||||||
| 1046 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1047 | rewriter.eraseOp(op); | ||||||
| 1048 | return mlir::success(); | ||||||
| 1049 | } | ||||||
| 1050 | }; | ||||||
| 1051 | |||||||
| 1052 | /// Lower `fir.global_len` operation. | ||||||
| 1053 | struct GlobalLenOpConversion : public FIROpConversion<fir::GlobalLenOp> { | ||||||
| 1054 | using FIROpConversion::FIROpConversion; | ||||||
| 1055 | |||||||
| 1056 | mlir::LogicalResult | ||||||
| 1057 | matchAndRewrite(fir::GlobalLenOp globalLen, OpAdaptor adaptor, | ||||||
| 1058 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1059 | TODO(globalLen.getLoc(), "fir.global_len codegen")do { fir::emitFatalError(globalLen.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1059" ": not yet implemented: ") + llvm::Twine("fir.global_len codegen" ), false); } while (false); | ||||||
| 1060 | return mlir::failure(); | ||||||
| 1061 | } | ||||||
| 1062 | }; | ||||||
| 1063 | |||||||
| 1064 | /// Lower fir.len_param_index | ||||||
| 1065 | struct LenParamIndexOpConversion | ||||||
| 1066 | : public FIROpConversion<fir::LenParamIndexOp> { | ||||||
| 1067 | using FIROpConversion::FIROpConversion; | ||||||
| 1068 | |||||||
| 1069 | // FIXME: this should be specialized by the runtime target | ||||||
| 1070 | mlir::LogicalResult | ||||||
| 1071 | matchAndRewrite(fir::LenParamIndexOp lenp, OpAdaptor, | ||||||
| 1072 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1073 | TODO(lenp.getLoc(), "fir.len_param_index codegen")do { fir::emitFatalError(lenp.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1073" ": not yet implemented: ") + llvm::Twine("fir.len_param_index codegen" ), false); } while (false); | ||||||
| 1074 | } | ||||||
| 1075 | }; | ||||||
| 1076 | |||||||
| 1077 | /// Convert `!fir.emboxchar<!fir.char<KIND, ?>, #n>` into a sequence of | ||||||
| 1078 | /// instructions that generate `!llvm.struct<(ptr<ik>, i64)>`. The 1st element | ||||||
| 1079 | /// in this struct is a pointer. Its type is determined from `KIND`. The 2nd | ||||||
| 1080 | /// element is the length of the character buffer (`#n`). | ||||||
| 1081 | struct EmboxCharOpConversion : public FIROpConversion<fir::EmboxCharOp> { | ||||||
| 1082 | using FIROpConversion::FIROpConversion; | ||||||
| 1083 | |||||||
| 1084 | mlir::LogicalResult | ||||||
| 1085 | matchAndRewrite(fir::EmboxCharOp emboxChar, OpAdaptor adaptor, | ||||||
| 1086 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1087 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 1088 | |||||||
| 1089 | mlir::Value charBuffer = operands[0]; | ||||||
| 1090 | mlir::Value charBufferLen = operands[1]; | ||||||
| 1091 | |||||||
| 1092 | mlir::Location loc = emboxChar.getLoc(); | ||||||
| 1093 | mlir::Type llvmStructTy = convertType(emboxChar.getType()); | ||||||
| 1094 | auto llvmStruct = rewriter.create<mlir::LLVM::UndefOp>(loc, llvmStructTy); | ||||||
| 1095 | |||||||
| 1096 | mlir::Type lenTy = | ||||||
| 1097 | llvmStructTy.cast<mlir::LLVM::LLVMStructType>().getBody()[1]; | ||||||
| 1098 | mlir::Value lenAfterCast = integerCast(loc, rewriter, lenTy, charBufferLen); | ||||||
| 1099 | |||||||
| 1100 | mlir::Type addrTy = | ||||||
| 1101 | llvmStructTy.cast<mlir::LLVM::LLVMStructType>().getBody()[0]; | ||||||
| 1102 | if (addrTy != charBuffer.getType()) | ||||||
| 1103 | charBuffer = | ||||||
| 1104 | rewriter.create<mlir::LLVM::BitcastOp>(loc, addrTy, charBuffer); | ||||||
| 1105 | |||||||
| 1106 | auto insertBufferOp = rewriter.create<mlir::LLVM::InsertValueOp>( | ||||||
| 1107 | loc, llvmStruct, charBuffer, 0); | ||||||
| 1108 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( | ||||||
| 1109 | emboxChar, insertBufferOp, lenAfterCast, 1); | ||||||
| 1110 | |||||||
| 1111 | return mlir::success(); | ||||||
| 1112 | } | ||||||
| 1113 | }; | ||||||
| 1114 | } // namespace | ||||||
| 1115 | |||||||
| 1116 | /// Return the LLVMFuncOp corresponding to the standard malloc call. | ||||||
| 1117 | static mlir::LLVM::LLVMFuncOp | ||||||
| 1118 | getMalloc(fir::AllocMemOp op, mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 1119 | auto module = op->getParentOfType<mlir::ModuleOp>(); | ||||||
| 1120 | if (mlir::LLVM::LLVMFuncOp mallocFunc = | ||||||
| 1121 | module.lookupSymbol<mlir::LLVM::LLVMFuncOp>("malloc")) | ||||||
| 1122 | return mallocFunc; | ||||||
| 1123 | mlir::OpBuilder moduleBuilder( | ||||||
| 1124 | op->getParentOfType<mlir::ModuleOp>().getBodyRegion()); | ||||||
| 1125 | auto indexType = mlir::IntegerType::get(op.getContext(), 64); | ||||||
| 1126 | return moduleBuilder.create<mlir::LLVM::LLVMFuncOp>( | ||||||
| 1127 | rewriter.getUnknownLoc(), "malloc", | ||||||
| 1128 | mlir::LLVM::LLVMFunctionType::get(getVoidPtrType(op.getContext()), | ||||||
| 1129 | indexType, | ||||||
| 1130 | /*isVarArg=*/false)); | ||||||
| 1131 | } | ||||||
| 1132 | |||||||
| 1133 | /// Helper function for generating the LLVM IR that computes the distance | ||||||
| 1134 | /// in bytes between adjacent elements pointed to by a pointer | ||||||
| 1135 | /// of type \p ptrTy. The result is returned as a value of \p idxTy integer | ||||||
| 1136 | /// type. | ||||||
| 1137 | static mlir::Value | ||||||
| 1138 | computeElementDistance(mlir::Location loc, mlir::Type ptrTy, mlir::Type idxTy, | ||||||
| 1139 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 1140 | // Note that we cannot use something like | ||||||
| 1141 | // mlir::LLVM::getPrimitiveTypeSizeInBits() for the element type here. For | ||||||
| 1142 | // example, it returns 10 bytes for mlir::Float80Type for targets where it | ||||||
| 1143 | // occupies 16 bytes. Proper solution is probably to use | ||||||
| 1144 | // mlir::DataLayout::getTypeABIAlignment(), but DataLayout is not being set | ||||||
| 1145 | // yet (see llvm-project#57230). For the time being use the '(intptr_t)((type | ||||||
| 1146 | // *)0 + 1)' trick for all types. The generated instructions are optimized | ||||||
| 1147 | // into constant by the first pass of InstCombine, so it should not be a | ||||||
| 1148 | // performance issue. | ||||||
| 1149 | auto nullPtr = rewriter.create<mlir::LLVM::NullOp>(loc, ptrTy); | ||||||
| 1150 | auto gep = rewriter.create<mlir::LLVM::GEPOp>( | ||||||
| 1151 | loc, ptrTy, nullPtr, llvm::ArrayRef<mlir::LLVM::GEPArg>{1}); | ||||||
| 1152 | return rewriter.create<mlir::LLVM::PtrToIntOp>(loc, idxTy, gep); | ||||||
| 1153 | } | ||||||
| 1154 | |||||||
| 1155 | /// Return value of the stride in bytes between adjacent elements | ||||||
| 1156 | /// of LLVM type \p llTy. The result is returned as a value of | ||||||
| 1157 | /// \p idxTy integer type. | ||||||
| 1158 | static mlir::Value | ||||||
| 1159 | genTypeStrideInBytes(mlir::Location loc, mlir::Type idxTy, | ||||||
| 1160 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1161 | mlir::Type llTy) { | ||||||
| 1162 | // Create a pointer type and use computeElementDistance(). | ||||||
| 1163 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(llTy); | ||||||
| 1164 | return computeElementDistance(loc, ptrTy, idxTy, rewriter); | ||||||
| 1165 | } | ||||||
| 1166 | |||||||
| 1167 | namespace { | ||||||
| 1168 | /// Lower a `fir.allocmem` instruction into `llvm.call @malloc` | ||||||
| 1169 | struct AllocMemOpConversion : public FIROpConversion<fir::AllocMemOp> { | ||||||
| 1170 | using FIROpConversion::FIROpConversion; | ||||||
| 1171 | |||||||
| 1172 | mlir::LogicalResult | ||||||
| 1173 | matchAndRewrite(fir::AllocMemOp heap, OpAdaptor adaptor, | ||||||
| 1174 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1175 | mlir::Type heapTy = heap.getType(); | ||||||
| 1176 | mlir::Type ty = convertType(heapTy); | ||||||
| 1177 | mlir::LLVM::LLVMFuncOp mallocFunc = getMalloc(heap, rewriter); | ||||||
| 1178 | mlir::Location loc = heap.getLoc(); | ||||||
| 1179 | auto ity = lowerTy().indexType(); | ||||||
| 1180 | mlir::Type dataTy = fir::unwrapRefType(heapTy); | ||||||
| 1181 | if (fir::isRecordWithTypeParameters(fir::unwrapSequenceType(dataTy))) | ||||||
| 1182 | TODO(loc, "fir.allocmem codegen of derived type with length parameters")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1182" ": not yet implemented: ") + llvm::Twine("fir.allocmem codegen of derived type with length parameters" ), false); } while (false); | ||||||
| 1183 | mlir::Value size = genTypeSizeInBytes(loc, ity, rewriter, ty); | ||||||
| 1184 | if (auto scaleSize = genAllocationScaleSize(heap, ity, rewriter)) | ||||||
| 1185 | size = rewriter.create<mlir::LLVM::MulOp>(loc, ity, size, scaleSize); | ||||||
| 1186 | for (mlir::Value opnd : adaptor.getOperands()) | ||||||
| 1187 | size = rewriter.create<mlir::LLVM::MulOp>( | ||||||
| 1188 | loc, ity, size, integerCast(loc, rewriter, ity, opnd)); | ||||||
| 1189 | heap->setAttr("callee", mlir::SymbolRefAttr::get(mallocFunc)); | ||||||
| 1190 | auto malloc = rewriter.create<mlir::LLVM::CallOp>( | ||||||
| 1191 | loc, ::getVoidPtrType(heap.getContext()), size, heap->getAttrs()); | ||||||
| 1192 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(heap, ty, | ||||||
| 1193 | malloc.getResult()); | ||||||
| 1194 | return mlir::success(); | ||||||
| 1195 | } | ||||||
| 1196 | |||||||
| 1197 | /// Compute the allocation size in bytes of the element type of | ||||||
| 1198 | /// \p llTy pointer type. The result is returned as a value of \p idxTy | ||||||
| 1199 | /// integer type. | ||||||
| 1200 | mlir::Value genTypeSizeInBytes(mlir::Location loc, mlir::Type idxTy, | ||||||
| 1201 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1202 | mlir::Type llTy) const { | ||||||
| 1203 | auto ptrTy = llTy.dyn_cast<mlir::LLVM::LLVMPointerType>(); | ||||||
| 1204 | return computeElementDistance(loc, ptrTy, idxTy, rewriter); | ||||||
| 1205 | } | ||||||
| 1206 | }; | ||||||
| 1207 | } // namespace | ||||||
| 1208 | |||||||
| 1209 | /// Return the LLVMFuncOp corresponding to the standard free call. | ||||||
| 1210 | static mlir::LLVM::LLVMFuncOp | ||||||
| 1211 | getFree(fir::FreeMemOp op, mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 1212 | auto module = op->getParentOfType<mlir::ModuleOp>(); | ||||||
| 1213 | if (mlir::LLVM::LLVMFuncOp freeFunc = | ||||||
| 1214 | module.lookupSymbol<mlir::LLVM::LLVMFuncOp>("free")) | ||||||
| 1215 | return freeFunc; | ||||||
| 1216 | mlir::OpBuilder moduleBuilder(module.getBodyRegion()); | ||||||
| 1217 | auto voidType = mlir::LLVM::LLVMVoidType::get(op.getContext()); | ||||||
| 1218 | return moduleBuilder.create<mlir::LLVM::LLVMFuncOp>( | ||||||
| 1219 | rewriter.getUnknownLoc(), "free", | ||||||
| 1220 | mlir::LLVM::LLVMFunctionType::get(voidType, | ||||||
| 1221 | getVoidPtrType(op.getContext()), | ||||||
| 1222 | /*isVarArg=*/false)); | ||||||
| 1223 | } | ||||||
| 1224 | |||||||
| 1225 | static unsigned getDimension(mlir::LLVM::LLVMArrayType ty) { | ||||||
| 1226 | unsigned result = 1; | ||||||
| 1227 | for (auto eleTy = ty.getElementType().dyn_cast<mlir::LLVM::LLVMArrayType>(); | ||||||
| 1228 | eleTy; | ||||||
| 1229 | eleTy = eleTy.getElementType().dyn_cast<mlir::LLVM::LLVMArrayType>()) | ||||||
| 1230 | ++result; | ||||||
| 1231 | return result; | ||||||
| 1232 | } | ||||||
| 1233 | |||||||
| 1234 | namespace { | ||||||
| 1235 | /// Lower a `fir.freemem` instruction into `llvm.call @free` | ||||||
| 1236 | struct FreeMemOpConversion : public FIROpConversion<fir::FreeMemOp> { | ||||||
| 1237 | using FIROpConversion::FIROpConversion; | ||||||
| 1238 | |||||||
| 1239 | mlir::LogicalResult | ||||||
| 1240 | matchAndRewrite(fir::FreeMemOp freemem, OpAdaptor adaptor, | ||||||
| 1241 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1242 | mlir::LLVM::LLVMFuncOp freeFunc = getFree(freemem, rewriter); | ||||||
| 1243 | mlir::Location loc = freemem.getLoc(); | ||||||
| 1244 | auto bitcast = rewriter.create<mlir::LLVM::BitcastOp>( | ||||||
| 1245 | freemem.getLoc(), voidPtrTy(), adaptor.getOperands()[0]); | ||||||
| 1246 | freemem->setAttr("callee", mlir::SymbolRefAttr::get(freeFunc)); | ||||||
| 1247 | rewriter.create<mlir::LLVM::CallOp>( | ||||||
| 1248 | loc, mlir::TypeRange{}, mlir::ValueRange{bitcast}, freemem->getAttrs()); | ||||||
| 1249 | rewriter.eraseOp(freemem); | ||||||
| 1250 | return mlir::success(); | ||||||
| 1251 | } | ||||||
| 1252 | }; | ||||||
| 1253 | } // namespace | ||||||
| 1254 | |||||||
| 1255 | /// Common base class for embox to descriptor conversion. | ||||||
| 1256 | template <typename OP> | ||||||
| 1257 | struct EmboxCommonConversion : public FIROpConversion<OP> { | ||||||
| 1258 | using FIROpConversion<OP>::FIROpConversion; | ||||||
| 1259 | |||||||
| 1260 | static int getCFIAttr(fir::BaseBoxType boxTy) { | ||||||
| 1261 | auto eleTy = boxTy.getEleTy(); | ||||||
| 1262 | if (eleTy.isa<fir::PointerType>()) | ||||||
| 1263 | return CFI_attribute_pointer1; | ||||||
| 1264 | if (eleTy.isa<fir::HeapType>()) | ||||||
| 1265 | return CFI_attribute_allocatable2; | ||||||
| 1266 | return CFI_attribute_other0; | ||||||
| 1267 | } | ||||||
| 1268 | |||||||
| 1269 | // Get the element size and CFI type code of the boxed value. | ||||||
| 1270 | std::tuple<mlir::Value, mlir::Value> getSizeAndTypeCode( | ||||||
| 1271 | mlir::Location loc, mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1272 | mlir::Type boxEleTy, mlir::ValueRange lenParams = {}) const { | ||||||
| 1273 | auto i64Ty = mlir::IntegerType::get(rewriter.getContext(), 64); | ||||||
| 1274 | if (auto eleTy = fir::dyn_cast_ptrEleTy(boxEleTy)) | ||||||
| 1275 | boxEleTy = eleTy; | ||||||
| 1276 | if (auto seqTy = boxEleTy.dyn_cast<fir::SequenceType>()) | ||||||
| 1277 | return getSizeAndTypeCode(loc, rewriter, seqTy.getEleTy(), lenParams); | ||||||
| 1278 | if (boxEleTy.isa<mlir::NoneType>()) // unlimited polymorphic or assumed type | ||||||
| 1279 | return {rewriter.create<mlir::LLVM::ConstantOp>(loc, i64Ty, 0), | ||||||
| 1280 | this->genConstantOffset(loc, rewriter, CFI_type_other(-1))}; | ||||||
| 1281 | mlir::Value typeCodeVal = this->genConstantOffset( | ||||||
| 1282 | loc, rewriter, | ||||||
| 1283 | fir::getTypeCode(boxEleTy, this->lowerTy().getKindMap())); | ||||||
| 1284 | if (fir::isa_integer(boxEleTy) || boxEleTy.dyn_cast<fir::LogicalType>() || | ||||||
| 1285 | fir::isa_real(boxEleTy) || fir::isa_complex(boxEleTy)) | ||||||
| 1286 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, | ||||||
| 1287 | this->convertType(boxEleTy)), | ||||||
| 1288 | typeCodeVal}; | ||||||
| 1289 | if (auto charTy = boxEleTy.dyn_cast<fir::CharacterType>()) { | ||||||
| 1290 | mlir::Value size = | ||||||
| 1291 | genTypeStrideInBytes(loc, i64Ty, rewriter, this->convertType(charTy)); | ||||||
| 1292 | if (charTy.getLen() == fir::CharacterType::unknownLen()) { | ||||||
| 1293 | // Multiply the single character size by the length. | ||||||
| 1294 | assert(!lenParams.empty())(static_cast <bool> (!lenParams.empty()) ? void (0) : __assert_fail ("!lenParams.empty()", "flang/lib/Optimizer/CodeGen/CodeGen.cpp" , 1294, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1295 | auto len64 = FIROpConversion<OP>::integerCast(loc, rewriter, i64Ty, | ||||||
| 1296 | lenParams.back()); | ||||||
| 1297 | size = rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, size, len64); | ||||||
| 1298 | } | ||||||
| 1299 | return {size, typeCodeVal}; | ||||||
| 1300 | }; | ||||||
| 1301 | if (fir::isa_ref_type(boxEleTy)) { | ||||||
| 1302 | auto ptrTy = mlir::LLVM::LLVMPointerType::get( | ||||||
| 1303 | mlir::LLVM::LLVMVoidType::get(rewriter.getContext())); | ||||||
| 1304 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, ptrTy), typeCodeVal}; | ||||||
| 1305 | } | ||||||
| 1306 | if (boxEleTy.isa<fir::RecordType>()) | ||||||
| 1307 | return {genTypeStrideInBytes(loc, i64Ty, rewriter, | ||||||
| 1308 | this->convertType(boxEleTy)), | ||||||
| 1309 | typeCodeVal}; | ||||||
| 1310 | fir::emitFatalError(loc, "unhandled type in fir.box code generation"); | ||||||
| 1311 | } | ||||||
| 1312 | |||||||
| 1313 | /// Basic pattern to write a field in the descriptor | ||||||
| 1314 | mlir::Value insertField(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1315 | mlir::Location loc, mlir::Value dest, | ||||||
| 1316 | llvm::ArrayRef<std::int64_t> fldIndexes, | ||||||
| 1317 | mlir::Value value, bool bitcast = false) const { | ||||||
| 1318 | auto boxTy = dest.getType(); | ||||||
| 1319 | auto fldTy = this->getBoxEleTy(boxTy, fldIndexes); | ||||||
| 1320 | if (bitcast) | ||||||
| 1321 | value = rewriter.create<mlir::LLVM::BitcastOp>(loc, fldTy, value); | ||||||
| 1322 | else | ||||||
| 1323 | value = this->integerCast(loc, rewriter, fldTy, value); | ||||||
| 1324 | return rewriter.create<mlir::LLVM::InsertValueOp>(loc, dest, value, | ||||||
| 1325 | fldIndexes); | ||||||
| 1326 | } | ||||||
| 1327 | |||||||
| 1328 | inline mlir::Value | ||||||
| 1329 | insertBaseAddress(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1330 | mlir::Location loc, mlir::Value dest, | ||||||
| 1331 | mlir::Value base) const { | ||||||
| 1332 | return insertField(rewriter, loc, dest, {kAddrPosInBox}, base, | ||||||
| 1333 | /*bitCast=*/true); | ||||||
| 1334 | } | ||||||
| 1335 | |||||||
| 1336 | inline mlir::Value insertLowerBound(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1337 | mlir::Location loc, mlir::Value dest, | ||||||
| 1338 | unsigned dim, mlir::Value lb) const { | ||||||
| 1339 | return insertField(rewriter, loc, dest, | ||||||
| 1340 | {kDimsPosInBox, dim, kDimLowerBoundPos}, lb); | ||||||
| 1341 | } | ||||||
| 1342 | |||||||
| 1343 | inline mlir::Value insertExtent(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1344 | mlir::Location loc, mlir::Value dest, | ||||||
| 1345 | unsigned dim, mlir::Value extent) const { | ||||||
| 1346 | return insertField(rewriter, loc, dest, {kDimsPosInBox, dim, kDimExtentPos}, | ||||||
| 1347 | extent); | ||||||
| 1348 | } | ||||||
| 1349 | |||||||
| 1350 | inline mlir::Value insertStride(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1351 | mlir::Location loc, mlir::Value dest, | ||||||
| 1352 | unsigned dim, mlir::Value stride) const { | ||||||
| 1353 | return insertField(rewriter, loc, dest, {kDimsPosInBox, dim, kDimStridePos}, | ||||||
| 1354 | stride); | ||||||
| 1355 | } | ||||||
| 1356 | |||||||
| 1357 | /// Get the address of the type descriptor global variable that was created by | ||||||
| 1358 | /// lowering for derived type \p recType. | ||||||
| 1359 | mlir::Value getTypeDescriptor(mlir::ModuleOp mod, | ||||||
| 1360 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1361 | mlir::Location loc, | ||||||
| 1362 | fir::RecordType recType) const { | ||||||
| 1363 | std::string name = | ||||||
| 1364 | fir::NameUniquer::getTypeDescriptorName(recType.getName()); | ||||||
| 1365 | if (auto global = mod.template lookupSymbol<fir::GlobalOp>(name)) { | ||||||
| 1366 | auto ty = mlir::LLVM::LLVMPointerType::get( | ||||||
| 1367 | this->lowerTy().convertType(global.getType())); | ||||||
| 1368 | return rewriter.create<mlir::LLVM::AddressOfOp>(loc, ty, | ||||||
| 1369 | global.getSymName()); | ||||||
| 1370 | } | ||||||
| 1371 | if (auto global = mod.template lookupSymbol<mlir::LLVM::GlobalOp>(name)) { | ||||||
| 1372 | // The global may have already been translated to LLVM. | ||||||
| 1373 | auto ty = mlir::LLVM::LLVMPointerType::get(global.getType()); | ||||||
| 1374 | return rewriter.create<mlir::LLVM::AddressOfOp>(loc, ty, | ||||||
| 1375 | global.getSymName()); | ||||||
| 1376 | } | ||||||
| 1377 | // Type info derived types do not have type descriptors since they are the | ||||||
| 1378 | // types defining type descriptors. | ||||||
| 1379 | if (!this->options.ignoreMissingTypeDescriptors && | ||||||
| 1380 | !fir::NameUniquer::belongsToModule( | ||||||
| 1381 | name, Fortran::semantics::typeInfoBuiltinModule)) | ||||||
| 1382 | fir::emitFatalError( | ||||||
| 1383 | loc, "runtime derived type info descriptor was not generated"); | ||||||
| 1384 | return rewriter.create<mlir::LLVM::NullOp>( | ||||||
| 1385 | loc, ::getVoidPtrType(mod.getContext())); | ||||||
| 1386 | } | ||||||
| 1387 | |||||||
| 1388 | mlir::Value populateDescriptor(mlir::Location loc, mlir::ModuleOp mod, | ||||||
| 1389 | fir::BaseBoxType boxTy, mlir::Type inputType, | ||||||
| 1390 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1391 | unsigned rank, mlir::Value eleSize, | ||||||
| 1392 | mlir::Value cfiTy, | ||||||
| 1393 | mlir::Value typeDesc) const { | ||||||
| 1394 | auto convTy = this->lowerTy().convertBoxType(boxTy, rank); | ||||||
| 1395 | auto llvmBoxPtrTy = convTy.template cast<mlir::LLVM::LLVMPointerType>(); | ||||||
| 1396 | auto llvmBoxTy = llvmBoxPtrTy.getElementType(); | ||||||
| 1397 | bool isUnlimitedPolymorphic = fir::isUnlimitedPolymorphicType(boxTy); | ||||||
| 1398 | bool useInputType = fir::isPolymorphicType(boxTy) || isUnlimitedPolymorphic; | ||||||
| 1399 | mlir::Value descriptor = | ||||||
| 1400 | rewriter.create<mlir::LLVM::UndefOp>(loc, llvmBoxTy); | ||||||
| 1401 | descriptor = | ||||||
| 1402 | insertField(rewriter, loc, descriptor, {kElemLenPosInBox}, eleSize); | ||||||
| 1403 | descriptor = insertField(rewriter, loc, descriptor, {kVersionPosInBox}, | ||||||
| 1404 | this->genI32Constant(loc, rewriter, CFI_VERSION20180515)); | ||||||
| 1405 | descriptor = insertField(rewriter, loc, descriptor, {kRankPosInBox}, | ||||||
| 1406 | this->genI32Constant(loc, rewriter, rank)); | ||||||
| 1407 | descriptor = insertField(rewriter, loc, descriptor, {kTypePosInBox}, cfiTy); | ||||||
| 1408 | descriptor = | ||||||
| 1409 | insertField(rewriter, loc, descriptor, {kAttributePosInBox}, | ||||||
| 1410 | this->genI32Constant(loc, rewriter, getCFIAttr(boxTy))); | ||||||
| 1411 | const bool hasAddendum = fir::boxHasAddendum(boxTy); | ||||||
| 1412 | descriptor = | ||||||
| 1413 | insertField(rewriter, loc, descriptor, {kF18AddendumPosInBox}, | ||||||
| 1414 | this->genI32Constant(loc, rewriter, hasAddendum ? 1 : 0)); | ||||||
| 1415 | |||||||
| 1416 | if (hasAddendum) { | ||||||
| 1417 | unsigned typeDescFieldId = getTypeDescFieldId(boxTy); | ||||||
| 1418 | if (!typeDesc) { | ||||||
| 1419 | if (useInputType) { | ||||||
| 1420 | mlir::Type innerType = fir::unwrapInnerType(inputType); | ||||||
| 1421 | if (innerType && innerType.template isa<fir::RecordType>()) { | ||||||
| 1422 | auto recTy = innerType.template dyn_cast<fir::RecordType>(); | ||||||
| 1423 | typeDesc = getTypeDescriptor(mod, rewriter, loc, recTy); | ||||||
| 1424 | } else { | ||||||
| 1425 | // Unlimited polymorphic type descriptor with no record type. Set | ||||||
| 1426 | // type descriptor address to a clean state. | ||||||
| 1427 | typeDesc = rewriter.create<mlir::LLVM::NullOp>( | ||||||
| 1428 | loc, ::getVoidPtrType(mod.getContext())); | ||||||
| 1429 | } | ||||||
| 1430 | } else { | ||||||
| 1431 | typeDesc = getTypeDescriptor(mod, rewriter, loc, | ||||||
| 1432 | fir::unwrapIfDerived(boxTy)); | ||||||
| 1433 | } | ||||||
| 1434 | } | ||||||
| 1435 | if (typeDesc) | ||||||
| 1436 | descriptor = | ||||||
| 1437 | insertField(rewriter, loc, descriptor, {typeDescFieldId}, typeDesc, | ||||||
| 1438 | /*bitCast=*/true); | ||||||
| 1439 | } | ||||||
| 1440 | return descriptor; | ||||||
| 1441 | } | ||||||
| 1442 | |||||||
| 1443 | // Template used for fir::EmboxOp and fir::cg::XEmboxOp | ||||||
| 1444 | template <typename BOX> | ||||||
| 1445 | std::tuple<fir::BaseBoxType, mlir::Value, mlir::Value> | ||||||
| 1446 | consDescriptorPrefix(BOX box, mlir::Type inputType, | ||||||
| 1447 | mlir::ConversionPatternRewriter &rewriter, unsigned rank, | ||||||
| 1448 | [[maybe_unused]] mlir::ValueRange substrParams, | ||||||
| 1449 | mlir::ValueRange lenParams, mlir::Value sourceBox = {}, | ||||||
| 1450 | mlir::Type sourceBoxType = {}) const { | ||||||
| 1451 | auto loc = box.getLoc(); | ||||||
| 1452 | auto boxTy = box.getType().template dyn_cast<fir::BaseBoxType>(); | ||||||
| 1453 | bool useInputType = fir::isPolymorphicType(boxTy) && | ||||||
| 1454 | !fir::isUnlimitedPolymorphicType(inputType); | ||||||
| 1455 | llvm::SmallVector<mlir::Value> typeparams = lenParams; | ||||||
| 1456 | if constexpr (!std::is_same_v<BOX, fir::EmboxOp>) { | ||||||
| 1457 | if (!box.getSubstr().empty() && fir::hasDynamicSize(boxTy.getEleTy())) | ||||||
| 1458 | typeparams.push_back(substrParams[1]); | ||||||
| 1459 | } | ||||||
| 1460 | |||||||
| 1461 | // Write each of the fields with the appropriate values. | ||||||
| 1462 | // When emboxing an element to a polymorphic descriptor, use the | ||||||
| 1463 | // input type since the destination descriptor type has not the exact | ||||||
| 1464 | // information. | ||||||
| 1465 | auto [eleSize, cfiTy] = getSizeAndTypeCode( | ||||||
| 1466 | loc, rewriter, useInputType ? inputType : boxTy.getEleTy(), typeparams); | ||||||
| 1467 | |||||||
| 1468 | mlir::Value typeDesc; | ||||||
| 1469 | // When emboxing to a polymorphic box, get the type descriptor, type code | ||||||
| 1470 | // and element size from the source box if any. | ||||||
| 1471 | if (fir::isPolymorphicType(boxTy) && sourceBox) { | ||||||
| 1472 | typeDesc = | ||||||
| 1473 | this->loadTypeDescAddress(loc, sourceBoxType, sourceBox, rewriter); | ||||||
| 1474 | mlir::Type idxTy = this->lowerTy().indexType(); | ||||||
| 1475 | eleSize = this->getElementSizeFromBox(loc, idxTy, sourceBoxType, | ||||||
| 1476 | sourceBox, rewriter); | ||||||
| 1477 | cfiTy = this->getValueFromBox(loc, sourceBoxType, sourceBox, | ||||||
| 1478 | cfiTy.getType(), rewriter, kTypePosInBox); | ||||||
| 1479 | } | ||||||
| 1480 | auto mod = box->template getParentOfType<mlir::ModuleOp>(); | ||||||
| 1481 | mlir::Value descriptor = populateDescriptor( | ||||||
| 1482 | loc, mod, boxTy, inputType, rewriter, rank, eleSize, cfiTy, typeDesc); | ||||||
| 1483 | |||||||
| 1484 | return {boxTy, descriptor, eleSize}; | ||||||
| 1485 | } | ||||||
| 1486 | |||||||
| 1487 | std::tuple<fir::BaseBoxType, mlir::Value, mlir::Value> | ||||||
| 1488 | consDescriptorPrefix(fir::cg::XReboxOp box, mlir::Value loweredBox, | ||||||
| 1489 | mlir::ConversionPatternRewriter &rewriter, unsigned rank, | ||||||
| 1490 | mlir::ValueRange substrParams, | ||||||
| 1491 | mlir::ValueRange lenParams, | ||||||
| 1492 | mlir::Value typeDesc = {}) const { | ||||||
| 1493 | auto loc = box.getLoc(); | ||||||
| 1494 | auto boxTy = box.getType().dyn_cast<fir::BaseBoxType>(); | ||||||
| 1495 | auto inputBoxTy = box.getBox().getType().dyn_cast<fir::BaseBoxType>(); | ||||||
| 1496 | llvm::SmallVector<mlir::Value> typeparams = lenParams; | ||||||
| 1497 | if (!box.getSubstr().empty() && fir::hasDynamicSize(boxTy.getEleTy())) | ||||||
| 1498 | typeparams.push_back(substrParams[1]); | ||||||
| 1499 | |||||||
| 1500 | auto [eleSize, cfiTy] = | ||||||
| 1501 | getSizeAndTypeCode(loc, rewriter, boxTy.getEleTy(), typeparams); | ||||||
| 1502 | |||||||
| 1503 | // Reboxing to a polymorphic entity. eleSize and type code need to | ||||||
| 1504 | // be retrieved from the initial box and propagated to the new box. | ||||||
| 1505 | // If the initial box has an addendum, the type desc must be propagated as | ||||||
| 1506 | // well. | ||||||
| 1507 | if (fir::isPolymorphicType(boxTy)) { | ||||||
| 1508 | mlir::Type idxTy = this->lowerTy().indexType(); | ||||||
| 1509 | eleSize = | ||||||
| 1510 | this->getElementSizeFromBox(loc, idxTy, boxTy, loweredBox, rewriter); | ||||||
| 1511 | cfiTy = this->getValueFromBox(loc, boxTy, loweredBox, cfiTy.getType(), | ||||||
| 1512 | rewriter, kTypePosInBox); | ||||||
| 1513 | // TODO: For initial box that are unlimited polymorphic entities, this | ||||||
| 1514 | // code must be made conditional because unlimited polymorphic entities | ||||||
| 1515 | // with intrinsic type spec does not have addendum. | ||||||
| 1516 | if (fir::boxHasAddendum(inputBoxTy)) | ||||||
| 1517 | typeDesc = this->loadTypeDescAddress(loc, box.getBox().getType(), | ||||||
| 1518 | loweredBox, rewriter); | ||||||
| 1519 | } | ||||||
| 1520 | |||||||
| 1521 | auto mod = box->template getParentOfType<mlir::ModuleOp>(); | ||||||
| 1522 | mlir::Value descriptor = | ||||||
| 1523 | populateDescriptor(loc, mod, boxTy, box.getBox().getType(), rewriter, | ||||||
| 1524 | rank, eleSize, cfiTy, typeDesc); | ||||||
| 1525 | |||||||
| 1526 | return {boxTy, descriptor, eleSize}; | ||||||
| 1527 | } | ||||||
| 1528 | |||||||
| 1529 | // Compute the base address of a fir.box given the indices from the slice. | ||||||
| 1530 | // The indices from the "outer" dimensions (every dimension after the first | ||||||
| 1531 | // one (inlcuded) that is not a compile time constant) must have been | ||||||
| 1532 | // multiplied with the related extents and added together into \p outerOffset. | ||||||
| 1533 | mlir::Value | ||||||
| 1534 | genBoxOffsetGep(mlir::ConversionPatternRewriter &rewriter, mlir::Location loc, | ||||||
| 1535 | mlir::Value base, mlir::Value outerOffset, | ||||||
| 1536 | mlir::ValueRange cstInteriorIndices, | ||||||
| 1537 | mlir::ValueRange componentIndices, | ||||||
| 1538 | std::optional<mlir::Value> substringOffset) const { | ||||||
| 1539 | llvm::SmallVector<mlir::LLVM::GEPArg> gepArgs{outerOffset}; | ||||||
| 1540 | mlir::Type resultTy = | ||||||
| 1541 | base.getType().cast<mlir::LLVM::LLVMPointerType>().getElementType(); | ||||||
| 1542 | // Fortran is column major, llvm GEP is row major: reverse the indices here. | ||||||
| 1543 | for (mlir::Value interiorIndex : llvm::reverse(cstInteriorIndices)) { | ||||||
| 1544 | auto arrayTy = resultTy.dyn_cast<mlir::LLVM::LLVMArrayType>(); | ||||||
| 1545 | if (!arrayTy) | ||||||
| 1546 | fir::emitFatalError( | ||||||
| 1547 | loc, | ||||||
| 1548 | "corrupted GEP generated being generated in fir.embox/fir.rebox"); | ||||||
| 1549 | resultTy = arrayTy.getElementType(); | ||||||
| 1550 | gepArgs.push_back(interiorIndex); | ||||||
| 1551 | } | ||||||
| 1552 | for (mlir::Value componentIndex : componentIndices) { | ||||||
| 1553 | // Component indices can be field index to select a component, or array | ||||||
| 1554 | // index, to select an element in an array component. | ||||||
| 1555 | if (auto structTy = resultTy.dyn_cast<mlir::LLVM::LLVMStructType>()) { | ||||||
| 1556 | std::int64_t cstIndex = getConstantIntValue(componentIndex); | ||||||
| 1557 | resultTy = structTy.getBody()[cstIndex]; | ||||||
| 1558 | } else if (auto arrayTy = | ||||||
| 1559 | resultTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) { | ||||||
| 1560 | resultTy = arrayTy.getElementType(); | ||||||
| 1561 | } else { | ||||||
| 1562 | fir::emitFatalError(loc, "corrupted component GEP generated being " | ||||||
| 1563 | "generated in fir.embox/fir.rebox"); | ||||||
| 1564 | } | ||||||
| 1565 | gepArgs.push_back(componentIndex); | ||||||
| 1566 | } | ||||||
| 1567 | if (substringOffset) { | ||||||
| 1568 | if (auto arrayTy = resultTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) { | ||||||
| 1569 | gepArgs.push_back(*substringOffset); | ||||||
| 1570 | resultTy = arrayTy.getElementType(); | ||||||
| 1571 | } else { | ||||||
| 1572 | // If the CHARACTER length is dynamic, the whole base type should have | ||||||
| 1573 | // degenerated to an llvm.ptr<i[width]>, and there should not be any | ||||||
| 1574 | // cstInteriorIndices/componentIndices. The substring offset can be | ||||||
| 1575 | // added to the outterOffset since it applies on the same LLVM type. | ||||||
| 1576 | if (gepArgs.size() != 1) | ||||||
| 1577 | fir::emitFatalError(loc, | ||||||
| 1578 | "corrupted substring GEP in fir.embox/fir.rebox"); | ||||||
| 1579 | mlir::Type outterOffsetTy = gepArgs[0].get<mlir::Value>().getType(); | ||||||
| 1580 | mlir::Value cast = | ||||||
| 1581 | this->integerCast(loc, rewriter, outterOffsetTy, *substringOffset); | ||||||
| 1582 | |||||||
| 1583 | gepArgs[0] = rewriter.create<mlir::LLVM::AddOp>( | ||||||
| 1584 | loc, outterOffsetTy, gepArgs[0].get<mlir::Value>(), cast); | ||||||
| 1585 | } | ||||||
| 1586 | } | ||||||
| 1587 | resultTy = mlir::LLVM::LLVMPointerType::get(resultTy); | ||||||
| 1588 | return rewriter.create<mlir::LLVM::GEPOp>(loc, resultTy, base, gepArgs); | ||||||
| 1589 | } | ||||||
| 1590 | |||||||
| 1591 | template <typename BOX> | ||||||
| 1592 | void | ||||||
| 1593 | getSubcomponentIndices(BOX xbox, mlir::Value memref, | ||||||
| 1594 | mlir::ValueRange operands, | ||||||
| 1595 | mlir::SmallVectorImpl<mlir::Value> &indices) const { | ||||||
| 1596 | // For each field in the path add the offset to base via the args list. | ||||||
| 1597 | // In the most general case, some offsets must be computed since | ||||||
| 1598 | // they are not be known until runtime. | ||||||
| 1599 | if (fir::hasDynamicSize(fir::unwrapSequenceType( | ||||||
| 1600 | fir::unwrapPassByRefType(memref.getType())))) | ||||||
| 1601 | TODO(xbox.getLoc(),do { fir::emitFatalError(xbox.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1602" ": not yet implemented: ") + llvm::Twine("fir.embox codegen dynamic size component in derived type" ), false); } while (false) | ||||||
| 1602 | "fir.embox codegen dynamic size component in derived type")do { fir::emitFatalError(xbox.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1602" ": not yet implemented: ") + llvm::Twine("fir.embox codegen dynamic size component in derived type" ), false); } while (false); | ||||||
| 1603 | indices.append(operands.begin() + xbox.subcomponentOffset(), | ||||||
| 1604 | operands.begin() + xbox.subcomponentOffset() + | ||||||
| 1605 | xbox.getSubcomponent().size()); | ||||||
| 1606 | } | ||||||
| 1607 | |||||||
| 1608 | static bool isInGlobalOp(mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 1609 | auto *thisBlock = rewriter.getInsertionBlock(); | ||||||
| 1610 | return thisBlock && | ||||||
| 1611 | mlir::isa<mlir::LLVM::GlobalOp>(thisBlock->getParentOp()); | ||||||
| 1612 | } | ||||||
| 1613 | |||||||
| 1614 | /// If the embox is not in a globalOp body, allocate storage for the box; | ||||||
| 1615 | /// store the value inside and return the generated alloca. Return the input | ||||||
| 1616 | /// value otherwise. | ||||||
| 1617 | mlir::Value | ||||||
| 1618 | placeInMemoryIfNotGlobalInit(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1619 | mlir::Location loc, mlir::Type boxTy, | ||||||
| 1620 | mlir::Value boxValue) const { | ||||||
| 1621 | if (isInGlobalOp(rewriter)) | ||||||
| 1622 | return boxValue; | ||||||
| 1623 | auto boxPtrTy = mlir::LLVM::LLVMPointerType::get(boxValue.getType()); | ||||||
| 1624 | auto alloca = | ||||||
| 1625 | this->genAllocaWithType(loc, boxPtrTy, defaultAlign, rewriter); | ||||||
| 1626 | auto storeOp = rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, alloca); | ||||||
| 1627 | this->attachTBAATag(storeOp, boxTy, boxTy, nullptr); | ||||||
| 1628 | return alloca; | ||||||
| 1629 | } | ||||||
| 1630 | }; | ||||||
| 1631 | |||||||
| 1632 | /// Compute the extent of a triplet slice (lb:ub:step). | ||||||
| 1633 | static mlir::Value | ||||||
| 1634 | computeTripletExtent(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 1635 | mlir::Location loc, mlir::Value lb, mlir::Value ub, | ||||||
| 1636 | mlir::Value step, mlir::Value zero, mlir::Type type) { | ||||||
| 1637 | mlir::Value extent = rewriter.create<mlir::LLVM::SubOp>(loc, type, ub, lb); | ||||||
| 1638 | extent = rewriter.create<mlir::LLVM::AddOp>(loc, type, extent, step); | ||||||
| 1639 | extent = rewriter.create<mlir::LLVM::SDivOp>(loc, type, extent, step); | ||||||
| 1640 | // If the resulting extent is negative (`ub-lb` and `step` have different | ||||||
| 1641 | // signs), zero must be returned instead. | ||||||
| 1642 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 1643 | loc, mlir::LLVM::ICmpPredicate::sgt, extent, zero); | ||||||
| 1644 | return rewriter.create<mlir::LLVM::SelectOp>(loc, cmp, extent, zero); | ||||||
| 1645 | } | ||||||
| 1646 | |||||||
| 1647 | /// Create a generic box on a memory reference. This conversions lowers the | ||||||
| 1648 | /// abstract box to the appropriate, initialized descriptor. | ||||||
| 1649 | struct EmboxOpConversion : public EmboxCommonConversion<fir::EmboxOp> { | ||||||
| 1650 | using EmboxCommonConversion::EmboxCommonConversion; | ||||||
| 1651 | |||||||
| 1652 | mlir::LogicalResult | ||||||
| 1653 | matchAndRewrite(fir::EmboxOp embox, OpAdaptor adaptor, | ||||||
| 1654 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1655 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 1656 | mlir::Value sourceBox; | ||||||
| 1657 | mlir::Type sourceBoxType; | ||||||
| 1658 | if (embox.getSourceBox()) { | ||||||
| 1659 | sourceBox = operands[embox.getSourceBoxOffset()]; | ||||||
| 1660 | sourceBoxType = embox.getSourceBox().getType(); | ||||||
| 1661 | } | ||||||
| 1662 | assert(!embox.getShape() && "There should be no dims on this embox op")(static_cast <bool> (!embox.getShape() && "There should be no dims on this embox op" ) ? void (0) : __assert_fail ("!embox.getShape() && \"There should be no dims on this embox op\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 1662, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1663 | auto [boxTy, dest, eleSize] = consDescriptorPrefix( | ||||||
| 1664 | embox, fir::unwrapRefType(embox.getMemref().getType()), rewriter, | ||||||
| 1665 | /*rank=*/0, /*substrParams=*/mlir::ValueRange{}, | ||||||
| 1666 | adaptor.getTypeparams(), sourceBox, sourceBoxType); | ||||||
| 1667 | dest = insertBaseAddress(rewriter, embox.getLoc(), dest, operands[0]); | ||||||
| 1668 | if (fir::isDerivedTypeWithLenParams(boxTy)) { | ||||||
| 1669 | TODO(embox.getLoc(),do { fir::emitFatalError(embox.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1670" ": not yet implemented: ") + llvm::Twine("fir.embox codegen of derived with length parameters" ), false); } while (false) | ||||||
| 1670 | "fir.embox codegen of derived with length parameters")do { fir::emitFatalError(embox.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1670" ": not yet implemented: ") + llvm::Twine("fir.embox codegen of derived with length parameters" ), false); } while (false); | ||||||
| 1671 | return mlir::failure(); | ||||||
| 1672 | } | ||||||
| 1673 | auto result = | ||||||
| 1674 | placeInMemoryIfNotGlobalInit(rewriter, embox.getLoc(), boxTy, dest); | ||||||
| 1675 | rewriter.replaceOp(embox, result); | ||||||
| 1676 | return mlir::success(); | ||||||
| 1677 | } | ||||||
| 1678 | }; | ||||||
| 1679 | |||||||
| 1680 | /// Create a generic box on a memory reference. | ||||||
| 1681 | struct XEmboxOpConversion : public EmboxCommonConversion<fir::cg::XEmboxOp> { | ||||||
| 1682 | using EmboxCommonConversion::EmboxCommonConversion; | ||||||
| 1683 | |||||||
| 1684 | mlir::LogicalResult | ||||||
| 1685 | matchAndRewrite(fir::cg::XEmboxOp xbox, OpAdaptor adaptor, | ||||||
| 1686 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1687 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 1688 | mlir::Value sourceBox; | ||||||
| 1689 | mlir::Type sourceBoxType; | ||||||
| 1690 | if (xbox.getSourceBox()) { | ||||||
| 1691 | sourceBox = operands[xbox.getSourceBoxOffset()]; | ||||||
| 1692 | sourceBoxType = xbox.getSourceBox().getType(); | ||||||
| 1693 | } | ||||||
| 1694 | auto [boxTy, dest, eleSize] = consDescriptorPrefix( | ||||||
| 1695 | xbox, fir::unwrapRefType(xbox.getMemref().getType()), rewriter, | ||||||
| 1696 | xbox.getOutRank(), adaptor.getSubstr(), adaptor.getLenParams(), | ||||||
| 1697 | sourceBox, sourceBoxType); | ||||||
| 1698 | // Generate the triples in the dims field of the descriptor | ||||||
| 1699 | auto i64Ty = mlir::IntegerType::get(xbox.getContext(), 64); | ||||||
| 1700 | mlir::Value base = operands[0]; | ||||||
| 1701 | assert(!xbox.getShape().empty() && "must have a shape")(static_cast <bool> (!xbox.getShape().empty() && "must have a shape") ? void (0) : __assert_fail ("!xbox.getShape().empty() && \"must have a shape\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 1701, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1702 | unsigned shapeOffset = xbox.shapeOffset(); | ||||||
| 1703 | bool hasShift = !xbox.getShift().empty(); | ||||||
| 1704 | unsigned shiftOffset = xbox.shiftOffset(); | ||||||
| 1705 | bool hasSlice = !xbox.getSlice().empty(); | ||||||
| 1706 | unsigned sliceOffset = xbox.sliceOffset(); | ||||||
| 1707 | mlir::Location loc = xbox.getLoc(); | ||||||
| 1708 | mlir::Value zero = genConstantIndex(loc, i64Ty, rewriter, 0); | ||||||
| 1709 | mlir::Value one = genConstantIndex(loc, i64Ty, rewriter, 1); | ||||||
| 1710 | mlir::Value prevPtrOff = one; | ||||||
| 1711 | mlir::Type eleTy = boxTy.getEleTy(); | ||||||
| 1712 | const unsigned rank = xbox.getRank(); | ||||||
| 1713 | llvm::SmallVector<mlir::Value> cstInteriorIndices; | ||||||
| 1714 | unsigned constRows = 0; | ||||||
| 1715 | mlir::Value ptrOffset = zero; | ||||||
| 1716 | mlir::Type memEleTy = fir::dyn_cast_ptrEleTy(xbox.getMemref().getType()); | ||||||
| 1717 | assert(memEleTy.isa<fir::SequenceType>())(static_cast <bool> (memEleTy.isa<fir::SequenceType> ()) ? void (0) : __assert_fail ("memEleTy.isa<fir::SequenceType>()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 1717, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1718 | auto seqTy = memEleTy.cast<fir::SequenceType>(); | ||||||
| 1719 | mlir::Type seqEleTy = seqTy.getEleTy(); | ||||||
| 1720 | // Adjust the element scaling factor if the element is a dependent type. | ||||||
| 1721 | if (fir::hasDynamicSize(seqEleTy)) { | ||||||
| 1722 | if (auto charTy = seqEleTy.dyn_cast<fir::CharacterType>()) { | ||||||
| 1723 | prevPtrOff = eleSize; | ||||||
| 1724 | } else if (seqEleTy.isa<fir::RecordType>()) { | ||||||
| 1725 | // prevPtrOff = ; | ||||||
| 1726 | TODO(loc, "generate call to calculate size of PDT")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1726" ": not yet implemented: ") + llvm::Twine("generate call to calculate size of PDT" ), false); } while (false); | ||||||
| 1727 | } else { | ||||||
| 1728 | fir::emitFatalError(loc, "unexpected dynamic type"); | ||||||
| 1729 | } | ||||||
| 1730 | } else { | ||||||
| 1731 | constRows = seqTy.getConstantRows(); | ||||||
| 1732 | } | ||||||
| 1733 | |||||||
| 1734 | const auto hasSubcomp = !xbox.getSubcomponent().empty(); | ||||||
| 1735 | const bool hasSubstr = !xbox.getSubstr().empty(); | ||||||
| 1736 | // Initial element stride that will be use to compute the step in | ||||||
| 1737 | // each dimension. | ||||||
| 1738 | mlir::Value prevDimByteStride = eleSize; | ||||||
| 1739 | if (hasSubcomp) { | ||||||
| 1740 | // We have a subcomponent. The step value needs to be the number of | ||||||
| 1741 | // bytes per element (which is a derived type). | ||||||
| 1742 | prevDimByteStride = | ||||||
| 1743 | genTypeStrideInBytes(loc, i64Ty, rewriter, convertType(seqEleTy)); | ||||||
| 1744 | } else if (hasSubstr) { | ||||||
| 1745 | // We have a substring. The step value needs to be the number of bytes | ||||||
| 1746 | // per CHARACTER element. | ||||||
| 1747 | auto charTy = seqEleTy.cast<fir::CharacterType>(); | ||||||
| 1748 | if (fir::hasDynamicSize(charTy)) { | ||||||
| 1749 | prevDimByteStride = prevPtrOff; | ||||||
| 1750 | } else { | ||||||
| 1751 | prevDimByteStride = genConstantIndex( | ||||||
| 1752 | loc, i64Ty, rewriter, | ||||||
| 1753 | charTy.getLen() * lowerTy().characterBitsize(charTy) / 8); | ||||||
| 1754 | } | ||||||
| 1755 | } | ||||||
| 1756 | |||||||
| 1757 | // Process the array subspace arguments (shape, shift, etc.), if any, | ||||||
| 1758 | // translating everything to values in the descriptor wherever the entity | ||||||
| 1759 | // has a dynamic array dimension. | ||||||
| 1760 | for (unsigned di = 0, descIdx = 0; di < rank; ++di) { | ||||||
| 1761 | mlir::Value extent = operands[shapeOffset]; | ||||||
| 1762 | mlir::Value outerExtent = extent; | ||||||
| 1763 | bool skipNext = false; | ||||||
| 1764 | if (hasSlice) { | ||||||
| 1765 | mlir::Value off = operands[sliceOffset]; | ||||||
| 1766 | mlir::Value adj = one; | ||||||
| 1767 | if (hasShift) | ||||||
| 1768 | adj = operands[shiftOffset]; | ||||||
| 1769 | auto ao = rewriter.create<mlir::LLVM::SubOp>(loc, i64Ty, off, adj); | ||||||
| 1770 | if (constRows > 0) { | ||||||
| 1771 | cstInteriorIndices.push_back(ao); | ||||||
| 1772 | } else { | ||||||
| 1773 | auto dimOff = | ||||||
| 1774 | rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, ao, prevPtrOff); | ||||||
| 1775 | ptrOffset = | ||||||
| 1776 | rewriter.create<mlir::LLVM::AddOp>(loc, i64Ty, dimOff, ptrOffset); | ||||||
| 1777 | } | ||||||
| 1778 | if (mlir::isa_and_nonnull<fir::UndefOp>( | ||||||
| 1779 | xbox.getSlice()[3 * di + 1].getDefiningOp())) { | ||||||
| 1780 | // This dimension contains a scalar expression in the array slice op. | ||||||
| 1781 | // The dimension is loop invariant, will be dropped, and will not | ||||||
| 1782 | // appear in the descriptor. | ||||||
| 1783 | skipNext = true; | ||||||
| 1784 | } | ||||||
| 1785 | } | ||||||
| 1786 | if (!skipNext) { | ||||||
| 1787 | // store extent | ||||||
| 1788 | if (hasSlice) | ||||||
| 1789 | extent = computeTripletExtent(rewriter, loc, operands[sliceOffset], | ||||||
| 1790 | operands[sliceOffset + 1], | ||||||
| 1791 | operands[sliceOffset + 2], zero, i64Ty); | ||||||
| 1792 | // Lower bound is normalized to 0 for BIND(C) interoperability. | ||||||
| 1793 | mlir::Value lb = zero; | ||||||
| 1794 | const bool isaPointerOrAllocatable = | ||||||
| 1795 | eleTy.isa<fir::PointerType>() || eleTy.isa<fir::HeapType>(); | ||||||
| 1796 | // Lower bound is defaults to 1 for POINTER, ALLOCATABLE, and | ||||||
| 1797 | // denormalized descriptors. | ||||||
| 1798 | if (isaPointerOrAllocatable || !normalizedLowerBound(xbox)) | ||||||
| 1799 | lb = one; | ||||||
| 1800 | // If there is a shifted origin, and no fir.slice, and this is not | ||||||
| 1801 | // a normalized descriptor then use the value from the shift op as | ||||||
| 1802 | // the lower bound. | ||||||
| 1803 | if (hasShift && !(hasSlice || hasSubcomp || hasSubstr) && | ||||||
| 1804 | (isaPointerOrAllocatable || !normalizedLowerBound(xbox))) { | ||||||
| 1805 | lb = operands[shiftOffset]; | ||||||
| 1806 | auto extentIsEmpty = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 1807 | loc, mlir::LLVM::ICmpPredicate::eq, extent, zero); | ||||||
| 1808 | lb = rewriter.create<mlir::LLVM::SelectOp>(loc, extentIsEmpty, one, | ||||||
| 1809 | lb); | ||||||
| 1810 | } | ||||||
| 1811 | dest = insertLowerBound(rewriter, loc, dest, descIdx, lb); | ||||||
| 1812 | |||||||
| 1813 | dest = insertExtent(rewriter, loc, dest, descIdx, extent); | ||||||
| 1814 | |||||||
| 1815 | // store step (scaled by shaped extent) | ||||||
| 1816 | mlir::Value step = prevDimByteStride; | ||||||
| 1817 | if (hasSlice) | ||||||
| 1818 | step = rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, step, | ||||||
| 1819 | operands[sliceOffset + 2]); | ||||||
| 1820 | dest = insertStride(rewriter, loc, dest, descIdx, step); | ||||||
| 1821 | ++descIdx; | ||||||
| 1822 | } | ||||||
| 1823 | |||||||
| 1824 | // compute the stride and offset for the next natural dimension | ||||||
| 1825 | prevDimByteStride = rewriter.create<mlir::LLVM::MulOp>( | ||||||
| 1826 | loc, i64Ty, prevDimByteStride, outerExtent); | ||||||
| 1827 | if (constRows == 0) | ||||||
| 1828 | prevPtrOff = rewriter.create<mlir::LLVM::MulOp>(loc, i64Ty, prevPtrOff, | ||||||
| 1829 | outerExtent); | ||||||
| 1830 | else | ||||||
| 1831 | --constRows; | ||||||
| 1832 | |||||||
| 1833 | // increment iterators | ||||||
| 1834 | ++shapeOffset; | ||||||
| 1835 | if (hasShift) | ||||||
| 1836 | ++shiftOffset; | ||||||
| 1837 | if (hasSlice) | ||||||
| 1838 | sliceOffset += 3; | ||||||
| 1839 | } | ||||||
| 1840 | if (hasSlice || hasSubcomp || hasSubstr) { | ||||||
| 1841 | // Shift the base address. | ||||||
| 1842 | llvm::SmallVector<mlir::Value> fieldIndices; | ||||||
| 1843 | std::optional<mlir::Value> substringOffset; | ||||||
| 1844 | if (hasSubcomp) | ||||||
| 1845 | getSubcomponentIndices(xbox, xbox.getMemref(), operands, fieldIndices); | ||||||
| 1846 | if (hasSubstr) | ||||||
| 1847 | substringOffset = operands[xbox.substrOffset()]; | ||||||
| 1848 | base = genBoxOffsetGep(rewriter, loc, base, ptrOffset, cstInteriorIndices, | ||||||
| 1849 | fieldIndices, substringOffset); | ||||||
| 1850 | } | ||||||
| 1851 | dest = insertBaseAddress(rewriter, loc, dest, base); | ||||||
| 1852 | if (fir::isDerivedTypeWithLenParams(boxTy)) | ||||||
| 1853 | TODO(loc, "fir.embox codegen of derived with length parameters")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1853" ": not yet implemented: ") + llvm::Twine("fir.embox codegen of derived with length parameters" ), false); } while (false); | ||||||
| 1854 | |||||||
| 1855 | mlir::Value result = | ||||||
| 1856 | placeInMemoryIfNotGlobalInit(rewriter, loc, boxTy, dest); | ||||||
| 1857 | rewriter.replaceOp(xbox, result); | ||||||
| 1858 | return mlir::success(); | ||||||
| 1859 | } | ||||||
| 1860 | |||||||
| 1861 | /// Return true if `xbox` has a normalized lower bounds attribute. A box value | ||||||
| 1862 | /// that is neither a POINTER nor an ALLOCATABLE should be normalized to a | ||||||
| 1863 | /// zero origin lower bound for interoperability with BIND(C). | ||||||
| 1864 | inline static bool normalizedLowerBound(fir::cg::XEmboxOp xbox) { | ||||||
| 1865 | return xbox->hasAttr(fir::getNormalizedLowerBoundAttrName()); | ||||||
| 1866 | } | ||||||
| 1867 | }; | ||||||
| 1868 | |||||||
| 1869 | /// Create a new box given a box reference. | ||||||
| 1870 | struct XReboxOpConversion : public EmboxCommonConversion<fir::cg::XReboxOp> { | ||||||
| 1871 | using EmboxCommonConversion::EmboxCommonConversion; | ||||||
| 1872 | |||||||
| 1873 | mlir::LogicalResult | ||||||
| 1874 | matchAndRewrite(fir::cg::XReboxOp rebox, OpAdaptor adaptor, | ||||||
| 1875 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 1876 | mlir::Location loc = rebox.getLoc(); | ||||||
| 1877 | mlir::Type idxTy = lowerTy().indexType(); | ||||||
| 1878 | mlir::Value loweredBox = adaptor.getOperands()[0]; | ||||||
| 1879 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 1880 | |||||||
| 1881 | // Inside a fir.global, the input box was produced as an llvm.struct<> | ||||||
| 1882 | // because objects cannot be handled in memory inside a fir.global body that | ||||||
| 1883 | // must be constant foldable. However, the type translation are not | ||||||
| 1884 | // contextual, so the fir.box<T> type of the operation that produced the | ||||||
| 1885 | // fir.box was translated to an llvm.ptr<llvm.struct<>> and the MLIR pass | ||||||
| 1886 | // manager inserted a builtin.unrealized_conversion_cast that was inserted | ||||||
| 1887 | // and needs to be removed here. | ||||||
| 1888 | if (isInGlobalOp(rewriter)) | ||||||
| 1889 | if (auto unrealizedCast = | ||||||
| 1890 | loweredBox.getDefiningOp<mlir::UnrealizedConversionCastOp>()) | ||||||
| 1891 | loweredBox = unrealizedCast.getInputs()[0]; | ||||||
| 1892 | |||||||
| 1893 | // Create new descriptor and fill its non-shape related data. | ||||||
| 1894 | llvm::SmallVector<mlir::Value, 2> lenParams; | ||||||
| 1895 | mlir::Type inputEleTy = getInputEleTy(rebox); | ||||||
| 1896 | if (auto charTy = inputEleTy.dyn_cast<fir::CharacterType>()) { | ||||||
| 1897 | mlir::Value len = getElementSizeFromBox( | ||||||
| 1898 | loc, idxTy, rebox.getBox().getType(), loweredBox, rewriter); | ||||||
| 1899 | if (charTy.getFKind() != 1) { | ||||||
| 1900 | mlir::Value width = | ||||||
| 1901 | genConstantIndex(loc, idxTy, rewriter, charTy.getFKind()); | ||||||
| 1902 | len = rewriter.create<mlir::LLVM::SDivOp>(loc, idxTy, len, width); | ||||||
| 1903 | } | ||||||
| 1904 | lenParams.emplace_back(len); | ||||||
| 1905 | } else if (auto recTy = inputEleTy.dyn_cast<fir::RecordType>()) { | ||||||
| 1906 | if (recTy.getNumLenParams() != 0) | ||||||
| 1907 | TODO(loc, "reboxing descriptor of derived type with length parameters")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "1907" ": not yet implemented: ") + llvm::Twine("reboxing descriptor of derived type with length parameters" ), false); } while (false); | ||||||
| 1908 | } | ||||||
| 1909 | |||||||
| 1910 | // Rebox on polymorphic entities needs to carry over the dynamic type. | ||||||
| 1911 | mlir::Value typeDescAddr; | ||||||
| 1912 | if (rebox.getBox().getType().isa<fir::ClassType>() && | ||||||
| 1913 | rebox.getType().isa<fir::ClassType>()) | ||||||
| 1914 | typeDescAddr = loadTypeDescAddress(loc, rebox.getBox().getType(), | ||||||
| 1915 | loweredBox, rewriter); | ||||||
| 1916 | |||||||
| 1917 | auto [boxTy, dest, eleSize] = | ||||||
| 1918 | consDescriptorPrefix(rebox, loweredBox, rewriter, rebox.getOutRank(), | ||||||
| 1919 | adaptor.getSubstr(), lenParams, typeDescAddr); | ||||||
| 1920 | |||||||
| 1921 | // Read input extents, strides, and base address | ||||||
| 1922 | llvm::SmallVector<mlir::Value> inputExtents; | ||||||
| 1923 | llvm::SmallVector<mlir::Value> inputStrides; | ||||||
| 1924 | const unsigned inputRank = rebox.getRank(); | ||||||
| 1925 | for (unsigned dim = 0; dim < inputRank; ++dim) { | ||||||
| 1926 | llvm::SmallVector<mlir::Value, 3> dimInfo = | ||||||
| 1927 | getDimsFromBox(loc, {idxTy, idxTy, idxTy}, rebox.getBox().getType(), | ||||||
| 1928 | loweredBox, dim, rewriter); | ||||||
| 1929 | inputExtents.emplace_back(dimInfo[1]); | ||||||
| 1930 | inputStrides.emplace_back(dimInfo[2]); | ||||||
| 1931 | } | ||||||
| 1932 | |||||||
| 1933 | mlir::Type baseTy = getBaseAddrTypeFromBox(loweredBox.getType()); | ||||||
| 1934 | mlir::Value baseAddr = getBaseAddrFromBox( | ||||||
| 1935 | loc, baseTy, rebox.getBox().getType(), loweredBox, rewriter); | ||||||
| 1936 | |||||||
| 1937 | if (!rebox.getSlice().empty() || !rebox.getSubcomponent().empty()) | ||||||
| 1938 | return sliceBox(rebox, boxTy, dest, baseAddr, inputExtents, inputStrides, | ||||||
| 1939 | operands, rewriter); | ||||||
| 1940 | return reshapeBox(rebox, boxTy, dest, baseAddr, inputExtents, inputStrides, | ||||||
| 1941 | operands, rewriter); | ||||||
| 1942 | } | ||||||
| 1943 | |||||||
| 1944 | private: | ||||||
| 1945 | /// Write resulting shape and base address in descriptor, and replace rebox | ||||||
| 1946 | /// op. | ||||||
| 1947 | mlir::LogicalResult | ||||||
| 1948 | finalizeRebox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, | ||||||
| 1949 | mlir::Value base, mlir::ValueRange lbounds, | ||||||
| 1950 | mlir::ValueRange extents, mlir::ValueRange strides, | ||||||
| 1951 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 1952 | mlir::Location loc = rebox.getLoc(); | ||||||
| 1953 | mlir::Value zero = | ||||||
| 1954 | genConstantIndex(loc, lowerTy().indexType(), rewriter, 0); | ||||||
| 1955 | mlir::Value one = genConstantIndex(loc, lowerTy().indexType(), rewriter, 1); | ||||||
| 1956 | for (auto iter : llvm::enumerate(llvm::zip(extents, strides))) { | ||||||
| 1957 | mlir::Value extent = std::get<0>(iter.value()); | ||||||
| 1958 | unsigned dim = iter.index(); | ||||||
| 1959 | mlir::Value lb = one; | ||||||
| 1960 | if (!lbounds.empty()) { | ||||||
| 1961 | lb = lbounds[dim]; | ||||||
| 1962 | auto extentIsEmpty = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 1963 | loc, mlir::LLVM::ICmpPredicate::eq, extent, zero); | ||||||
| 1964 | lb = rewriter.create<mlir::LLVM::SelectOp>(loc, extentIsEmpty, one, lb); | ||||||
| 1965 | }; | ||||||
| 1966 | dest = insertLowerBound(rewriter, loc, dest, dim, lb); | ||||||
| 1967 | dest = insertExtent(rewriter, loc, dest, dim, extent); | ||||||
| 1968 | dest = insertStride(rewriter, loc, dest, dim, std::get<1>(iter.value())); | ||||||
| 1969 | } | ||||||
| 1970 | dest = insertBaseAddress(rewriter, loc, dest, base); | ||||||
| 1971 | mlir::Value result = | ||||||
| 1972 | placeInMemoryIfNotGlobalInit(rewriter, rebox.getLoc(), destBoxTy, dest); | ||||||
| 1973 | rewriter.replaceOp(rebox, result); | ||||||
| 1974 | return mlir::success(); | ||||||
| 1975 | } | ||||||
| 1976 | |||||||
| 1977 | // Apply slice given the base address, extents and strides of the input box. | ||||||
| 1978 | mlir::LogicalResult | ||||||
| 1979 | sliceBox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, | ||||||
| 1980 | mlir::Value base, mlir::ValueRange inputExtents, | ||||||
| 1981 | mlir::ValueRange inputStrides, mlir::ValueRange operands, | ||||||
| 1982 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 1983 | mlir::Location loc = rebox.getLoc(); | ||||||
| 1984 | mlir::Type voidPtrTy = ::getVoidPtrType(rebox.getContext()); | ||||||
| 1985 | mlir::Type idxTy = lowerTy().indexType(); | ||||||
| 1986 | mlir::Value zero = genConstantIndex(loc, idxTy, rewriter, 0); | ||||||
| 1987 | // Apply subcomponent and substring shift on base address. | ||||||
| 1988 | if (!rebox.getSubcomponent().empty() || !rebox.getSubstr().empty()) { | ||||||
| 1989 | // Cast to inputEleTy* so that a GEP can be used. | ||||||
| 1990 | mlir::Type inputEleTy = getInputEleTy(rebox); | ||||||
| 1991 | auto llvmElePtrTy = | ||||||
| 1992 | mlir::LLVM::LLVMPointerType::get(convertType(inputEleTy)); | ||||||
| 1993 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, llvmElePtrTy, base); | ||||||
| 1994 | |||||||
| 1995 | llvm::SmallVector<mlir::Value> fieldIndices; | ||||||
| 1996 | std::optional<mlir::Value> substringOffset; | ||||||
| 1997 | if (!rebox.getSubcomponent().empty()) | ||||||
| 1998 | getSubcomponentIndices(rebox, rebox.getBox(), operands, fieldIndices); | ||||||
| 1999 | if (!rebox.getSubstr().empty()) | ||||||
| 2000 | substringOffset = operands[rebox.substrOffset()]; | ||||||
| 2001 | base = genBoxOffsetGep(rewriter, loc, base, zero, | ||||||
| 2002 | /*cstInteriorIndices=*/std::nullopt, fieldIndices, | ||||||
| 2003 | substringOffset); | ||||||
| 2004 | } | ||||||
| 2005 | |||||||
| 2006 | if (rebox.getSlice().empty()) | ||||||
| 2007 | // The array section is of the form array[%component][substring], keep | ||||||
| 2008 | // the input array extents and strides. | ||||||
| 2009 | return finalizeRebox(rebox, destBoxTy, dest, base, | ||||||
| 2010 | /*lbounds*/ std::nullopt, inputExtents, inputStrides, | ||||||
| 2011 | rewriter); | ||||||
| 2012 | |||||||
| 2013 | // Strides from the fir.box are in bytes. | ||||||
| 2014 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, voidPtrTy, base); | ||||||
| 2015 | |||||||
| 2016 | // The slice is of the form array(i:j:k)[%component]. Compute new extents | ||||||
| 2017 | // and strides. | ||||||
| 2018 | llvm::SmallVector<mlir::Value> slicedExtents; | ||||||
| 2019 | llvm::SmallVector<mlir::Value> slicedStrides; | ||||||
| 2020 | mlir::Value one = genConstantIndex(loc, idxTy, rewriter, 1); | ||||||
| 2021 | const bool sliceHasOrigins = !rebox.getShift().empty(); | ||||||
| 2022 | unsigned sliceOps = rebox.sliceOffset(); | ||||||
| 2023 | unsigned shiftOps = rebox.shiftOffset(); | ||||||
| 2024 | auto strideOps = inputStrides.begin(); | ||||||
| 2025 | const unsigned inputRank = inputStrides.size(); | ||||||
| 2026 | for (unsigned i = 0; i < inputRank; | ||||||
| 2027 | ++i, ++strideOps, ++shiftOps, sliceOps += 3) { | ||||||
| 2028 | mlir::Value sliceLb = | ||||||
| 2029 | integerCast(loc, rewriter, idxTy, operands[sliceOps]); | ||||||
| 2030 | mlir::Value inputStride = *strideOps; // already idxTy | ||||||
| 2031 | // Apply origin shift: base += (lb-shift)*input_stride | ||||||
| 2032 | mlir::Value sliceOrigin = | ||||||
| 2033 | sliceHasOrigins | ||||||
| 2034 | ? integerCast(loc, rewriter, idxTy, operands[shiftOps]) | ||||||
| 2035 | : one; | ||||||
| 2036 | mlir::Value diff = | ||||||
| 2037 | rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, sliceLb, sliceOrigin); | ||||||
| 2038 | mlir::Value offset = | ||||||
| 2039 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, inputStride); | ||||||
| 2040 | base = genGEP(loc, voidPtrTy, rewriter, base, offset); | ||||||
| 2041 | // Apply upper bound and step if this is a triplet. Otherwise, the | ||||||
| 2042 | // dimension is dropped and no extents/strides are computed. | ||||||
| 2043 | mlir::Value upper = operands[sliceOps + 1]; | ||||||
| 2044 | const bool isTripletSlice = | ||||||
| 2045 | !mlir::isa_and_nonnull<mlir::LLVM::UndefOp>(upper.getDefiningOp()); | ||||||
| 2046 | if (isTripletSlice) { | ||||||
| 2047 | mlir::Value step = | ||||||
| 2048 | integerCast(loc, rewriter, idxTy, operands[sliceOps + 2]); | ||||||
| 2049 | // extent = ub-lb+step/step | ||||||
| 2050 | mlir::Value sliceUb = integerCast(loc, rewriter, idxTy, upper); | ||||||
| 2051 | mlir::Value extent = computeTripletExtent(rewriter, loc, sliceLb, | ||||||
| 2052 | sliceUb, step, zero, idxTy); | ||||||
| 2053 | slicedExtents.emplace_back(extent); | ||||||
| 2054 | // stride = step*input_stride | ||||||
| 2055 | mlir::Value stride = | ||||||
| 2056 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, step, inputStride); | ||||||
| 2057 | slicedStrides.emplace_back(stride); | ||||||
| 2058 | } | ||||||
| 2059 | } | ||||||
| 2060 | return finalizeRebox(rebox, destBoxTy, dest, base, /*lbounds*/ std::nullopt, | ||||||
| 2061 | slicedExtents, slicedStrides, rewriter); | ||||||
| 2062 | } | ||||||
| 2063 | |||||||
| 2064 | /// Apply a new shape to the data described by a box given the base address, | ||||||
| 2065 | /// extents and strides of the box. | ||||||
| 2066 | mlir::LogicalResult | ||||||
| 2067 | reshapeBox(fir::cg::XReboxOp rebox, mlir::Type destBoxTy, mlir::Value dest, | ||||||
| 2068 | mlir::Value base, mlir::ValueRange inputExtents, | ||||||
| 2069 | mlir::ValueRange inputStrides, mlir::ValueRange operands, | ||||||
| 2070 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 2071 | mlir::ValueRange reboxShifts{operands.begin() + rebox.shiftOffset(), | ||||||
| 2072 | operands.begin() + rebox.shiftOffset() + | ||||||
| 2073 | rebox.getShift().size()}; | ||||||
| 2074 | if (rebox.getShape().empty()) { | ||||||
| 2075 | // Only setting new lower bounds. | ||||||
| 2076 | return finalizeRebox(rebox, destBoxTy, dest, base, reboxShifts, | ||||||
| 2077 | inputExtents, inputStrides, rewriter); | ||||||
| 2078 | } | ||||||
| 2079 | |||||||
| 2080 | mlir::Location loc = rebox.getLoc(); | ||||||
| 2081 | // Strides from the fir.box are in bytes. | ||||||
| 2082 | mlir::Type voidPtrTy = ::getVoidPtrType(rebox.getContext()); | ||||||
| 2083 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, voidPtrTy, base); | ||||||
| 2084 | |||||||
| 2085 | llvm::SmallVector<mlir::Value> newStrides; | ||||||
| 2086 | llvm::SmallVector<mlir::Value> newExtents; | ||||||
| 2087 | mlir::Type idxTy = lowerTy().indexType(); | ||||||
| 2088 | // First stride from input box is kept. The rest is assumed contiguous | ||||||
| 2089 | // (it is not possible to reshape otherwise). If the input is scalar, | ||||||
| 2090 | // which may be OK if all new extents are ones, the stride does not | ||||||
| 2091 | // matter, use one. | ||||||
| 2092 | mlir::Value stride = inputStrides.empty() | ||||||
| 2093 | ? genConstantIndex(loc, idxTy, rewriter, 1) | ||||||
| 2094 | : inputStrides[0]; | ||||||
| 2095 | for (unsigned i = 0; i < rebox.getShape().size(); ++i) { | ||||||
| 2096 | mlir::Value rawExtent = operands[rebox.shapeOffset() + i]; | ||||||
| 2097 | mlir::Value extent = integerCast(loc, rewriter, idxTy, rawExtent); | ||||||
| 2098 | newExtents.emplace_back(extent); | ||||||
| 2099 | newStrides.emplace_back(stride); | ||||||
| 2100 | // nextStride = extent * stride; | ||||||
| 2101 | stride = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, extent, stride); | ||||||
| 2102 | } | ||||||
| 2103 | return finalizeRebox(rebox, destBoxTy, dest, base, reboxShifts, newExtents, | ||||||
| 2104 | newStrides, rewriter); | ||||||
| 2105 | } | ||||||
| 2106 | |||||||
| 2107 | /// Return scalar element type of the input box. | ||||||
| 2108 | static mlir::Type getInputEleTy(fir::cg::XReboxOp rebox) { | ||||||
| 2109 | auto ty = fir::dyn_cast_ptrOrBoxEleTy(rebox.getBox().getType()); | ||||||
| 2110 | if (auto seqTy = ty.dyn_cast<fir::SequenceType>()) | ||||||
| 2111 | return seqTy.getEleTy(); | ||||||
| 2112 | return ty; | ||||||
| 2113 | } | ||||||
| 2114 | }; | ||||||
| 2115 | |||||||
| 2116 | /// Lower `fir.emboxproc` operation. Creates a procedure box. | ||||||
| 2117 | /// TODO: Part of supporting Fortran 2003 procedure pointers. | ||||||
| 2118 | struct EmboxProcOpConversion : public FIROpConversion<fir::EmboxProcOp> { | ||||||
| 2119 | using FIROpConversion::FIROpConversion; | ||||||
| 2120 | |||||||
| 2121 | mlir::LogicalResult | ||||||
| 2122 | matchAndRewrite(fir::EmboxProcOp emboxproc, OpAdaptor adaptor, | ||||||
| 2123 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2124 | TODO(emboxproc.getLoc(), "fir.emboxproc codegen")do { fir::emitFatalError(emboxproc.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2124" ": not yet implemented: ") + llvm::Twine("fir.emboxproc codegen" ), false); } while (false); | ||||||
| 2125 | return mlir::failure(); | ||||||
| 2126 | } | ||||||
| 2127 | }; | ||||||
| 2128 | |||||||
| 2129 | // Code shared between insert_value and extract_value Ops. | ||||||
| 2130 | struct ValueOpCommon { | ||||||
| 2131 | // Translate the arguments pertaining to any multidimensional array to | ||||||
| 2132 | // row-major order for LLVM-IR. | ||||||
| 2133 | static void toRowMajor(llvm::SmallVectorImpl<int64_t> &indices, | ||||||
| 2134 | mlir::Type ty) { | ||||||
| 2135 | assert(ty && "type is null")(static_cast <bool> (ty && "type is null") ? void (0) : __assert_fail ("ty && \"type is null\"", "flang/lib/Optimizer/CodeGen/CodeGen.cpp" , 2135, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2136 | const auto end = indices.size(); | ||||||
| 2137 | for (std::remove_const_t<decltype(end)> i = 0; i < end; ++i) { | ||||||
| 2138 | if (auto seq = ty.dyn_cast<mlir::LLVM::LLVMArrayType>()) { | ||||||
| 2139 | const auto dim = getDimension(seq); | ||||||
| 2140 | if (dim > 1) { | ||||||
| 2141 | auto ub = std::min(i + dim, end); | ||||||
| 2142 | std::reverse(indices.begin() + i, indices.begin() + ub); | ||||||
| 2143 | i += dim - 1; | ||||||
| 2144 | } | ||||||
| 2145 | ty = getArrayElementType(seq); | ||||||
| 2146 | } else if (auto st = ty.dyn_cast<mlir::LLVM::LLVMStructType>()) { | ||||||
| 2147 | ty = st.getBody()[indices[i]]; | ||||||
| 2148 | } else { | ||||||
| 2149 | llvm_unreachable("index into invalid type")::llvm::llvm_unreachable_internal("index into invalid type", "flang/lib/Optimizer/CodeGen/CodeGen.cpp" , 2149); | ||||||
| 2150 | } | ||||||
| 2151 | } | ||||||
| 2152 | } | ||||||
| 2153 | |||||||
| 2154 | static llvm::SmallVector<int64_t> | ||||||
| 2155 | collectIndices(mlir::ConversionPatternRewriter &rewriter, | ||||||
| 2156 | mlir::ArrayAttr arrAttr) { | ||||||
| 2157 | llvm::SmallVector<int64_t> indices; | ||||||
| 2158 | for (auto i = arrAttr.begin(), e = arrAttr.end(); i != e; ++i) { | ||||||
| 2159 | if (auto intAttr = i->dyn_cast<mlir::IntegerAttr>()) { | ||||||
| 2160 | indices.push_back(intAttr.getInt()); | ||||||
| 2161 | } else { | ||||||
| 2162 | auto fieldName = i->cast<mlir::StringAttr>().getValue(); | ||||||
| 2163 | ++i; | ||||||
| 2164 | auto ty = i->cast<mlir::TypeAttr>().getValue(); | ||||||
| 2165 | auto index = ty.cast<fir::RecordType>().getFieldIndex(fieldName); | ||||||
| 2166 | indices.push_back(index); | ||||||
| 2167 | } | ||||||
| 2168 | } | ||||||
| 2169 | return indices; | ||||||
| 2170 | } | ||||||
| 2171 | |||||||
| 2172 | private: | ||||||
| 2173 | static mlir::Type getArrayElementType(mlir::LLVM::LLVMArrayType ty) { | ||||||
| 2174 | auto eleTy = ty.getElementType(); | ||||||
| 2175 | while (auto arrTy = eleTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) | ||||||
| 2176 | eleTy = arrTy.getElementType(); | ||||||
| 2177 | return eleTy; | ||||||
| 2178 | } | ||||||
| 2179 | }; | ||||||
| 2180 | |||||||
| 2181 | namespace { | ||||||
| 2182 | /// Extract a subobject value from an ssa-value of aggregate type | ||||||
| 2183 | struct ExtractValueOpConversion | ||||||
| 2184 | : public FIROpAndTypeConversion<fir::ExtractValueOp>, | ||||||
| 2185 | public ValueOpCommon { | ||||||
| 2186 | using FIROpAndTypeConversion::FIROpAndTypeConversion; | ||||||
| 2187 | |||||||
| 2188 | mlir::LogicalResult | ||||||
| 2189 | doRewrite(fir::ExtractValueOp extractVal, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 2190 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2191 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 2192 | auto indices = collectIndices(rewriter, extractVal.getCoor()); | ||||||
| 2193 | toRowMajor(indices, operands[0].getType()); | ||||||
| 2194 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractValueOp>( | ||||||
| 2195 | extractVal, operands[0], indices); | ||||||
| 2196 | return mlir::success(); | ||||||
| 2197 | } | ||||||
| 2198 | }; | ||||||
| 2199 | |||||||
| 2200 | /// InsertValue is the generalized instruction for the composition of new | ||||||
| 2201 | /// aggregate type values. | ||||||
| 2202 | struct InsertValueOpConversion | ||||||
| 2203 | : public FIROpAndTypeConversion<fir::InsertValueOp>, | ||||||
| 2204 | public ValueOpCommon { | ||||||
| 2205 | using FIROpAndTypeConversion::FIROpAndTypeConversion; | ||||||
| 2206 | |||||||
| 2207 | mlir::LogicalResult | ||||||
| 2208 | doRewrite(fir::InsertValueOp insertVal, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 2209 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2210 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 2211 | auto indices = collectIndices(rewriter, insertVal.getCoor()); | ||||||
| 2212 | toRowMajor(indices, operands[0].getType()); | ||||||
| 2213 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( | ||||||
| 2214 | insertVal, operands[0], operands[1], indices); | ||||||
| 2215 | return mlir::success(); | ||||||
| 2216 | } | ||||||
| 2217 | }; | ||||||
| 2218 | |||||||
| 2219 | /// InsertOnRange inserts a value into a sequence over a range of offsets. | ||||||
| 2220 | struct InsertOnRangeOpConversion | ||||||
| 2221 | : public FIROpAndTypeConversion<fir::InsertOnRangeOp> { | ||||||
| 2222 | using FIROpAndTypeConversion::FIROpAndTypeConversion; | ||||||
| 2223 | |||||||
| 2224 | // Increments an array of subscripts in a row major fasion. | ||||||
| 2225 | void incrementSubscripts(llvm::ArrayRef<int64_t> dims, | ||||||
| 2226 | llvm::SmallVectorImpl<int64_t> &subscripts) const { | ||||||
| 2227 | for (size_t i = dims.size(); i > 0; --i) { | ||||||
| 2228 | if (++subscripts[i - 1] < dims[i - 1]) { | ||||||
| 2229 | return; | ||||||
| 2230 | } | ||||||
| 2231 | subscripts[i - 1] = 0; | ||||||
| 2232 | } | ||||||
| 2233 | } | ||||||
| 2234 | |||||||
| 2235 | mlir::LogicalResult | ||||||
| 2236 | doRewrite(fir::InsertOnRangeOp range, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 2237 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2238 | |||||||
| 2239 | llvm::SmallVector<std::int64_t> dims; | ||||||
| 2240 | auto type = adaptor.getOperands()[0].getType(); | ||||||
| 2241 | |||||||
| 2242 | // Iteratively extract the array dimensions from the type. | ||||||
| 2243 | while (auto t = type.dyn_cast<mlir::LLVM::LLVMArrayType>()) { | ||||||
| 2244 | dims.push_back(t.getNumElements()); | ||||||
| 2245 | type = t.getElementType(); | ||||||
| 2246 | } | ||||||
| 2247 | |||||||
| 2248 | llvm::SmallVector<std::int64_t> lBounds; | ||||||
| 2249 | llvm::SmallVector<std::int64_t> uBounds; | ||||||
| 2250 | |||||||
| 2251 | // Unzip the upper and lower bound and convert to a row major format. | ||||||
| 2252 | mlir::DenseIntElementsAttr coor = range.getCoor(); | ||||||
| 2253 | auto reversedCoor = llvm::reverse(coor.getValues<int64_t>()); | ||||||
| 2254 | for (auto i = reversedCoor.begin(), e = reversedCoor.end(); i != e; ++i) { | ||||||
| 2255 | uBounds.push_back(*i++); | ||||||
| 2256 | lBounds.push_back(*i); | ||||||
| 2257 | } | ||||||
| 2258 | |||||||
| 2259 | auto &subscripts = lBounds; | ||||||
| 2260 | auto loc = range.getLoc(); | ||||||
| 2261 | mlir::Value lastOp = adaptor.getOperands()[0]; | ||||||
| 2262 | mlir::Value insertVal = adaptor.getOperands()[1]; | ||||||
| 2263 | |||||||
| 2264 | while (subscripts != uBounds) { | ||||||
| 2265 | lastOp = rewriter.create<mlir::LLVM::InsertValueOp>( | ||||||
| 2266 | loc, lastOp, insertVal, subscripts); | ||||||
| 2267 | |||||||
| 2268 | incrementSubscripts(dims, subscripts); | ||||||
| 2269 | } | ||||||
| 2270 | |||||||
| 2271 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( | ||||||
| 2272 | range, lastOp, insertVal, subscripts); | ||||||
| 2273 | |||||||
| 2274 | return mlir::success(); | ||||||
| 2275 | } | ||||||
| 2276 | }; | ||||||
| 2277 | } // namespace | ||||||
| 2278 | |||||||
| 2279 | namespace { | ||||||
| 2280 | /// XArrayCoor is the address arithmetic on a dynamically shaped, sliced, | ||||||
| 2281 | /// shifted etc. array. | ||||||
| 2282 | /// (See the static restriction on coordinate_of.) array_coor determines the | ||||||
| 2283 | /// coordinate (location) of a specific element. | ||||||
| 2284 | struct XArrayCoorOpConversion | ||||||
| 2285 | : public FIROpAndTypeConversion<fir::cg::XArrayCoorOp> { | ||||||
| 2286 | using FIROpAndTypeConversion::FIROpAndTypeConversion; | ||||||
| 2287 | |||||||
| 2288 | mlir::LogicalResult | ||||||
| 2289 | doRewrite(fir::cg::XArrayCoorOp coor, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 2290 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2291 | auto loc = coor.getLoc(); | ||||||
| 2292 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 2293 | unsigned rank = coor.getRank(); | ||||||
| 2294 | assert(coor.getIndices().size() == rank)(static_cast <bool> (coor.getIndices().size() == rank) ? void (0) : __assert_fail ("coor.getIndices().size() == rank" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2294, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2295 | assert(coor.getShape().empty() || coor.getShape().size() == rank)(static_cast <bool> (coor.getShape().empty() || coor.getShape ().size() == rank) ? void (0) : __assert_fail ("coor.getShape().empty() || coor.getShape().size() == rank" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2295, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2296 | assert(coor.getShift().empty() || coor.getShift().size() == rank)(static_cast <bool> (coor.getShift().empty() || coor.getShift ().size() == rank) ? void (0) : __assert_fail ("coor.getShift().empty() || coor.getShift().size() == rank" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2296, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2297 | assert(coor.getSlice().empty() || coor.getSlice().size() == 3 * rank)(static_cast <bool> (coor.getSlice().empty() || coor.getSlice ().size() == 3 * rank) ? void (0) : __assert_fail ("coor.getSlice().empty() || coor.getSlice().size() == 3 * rank" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2297, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2298 | mlir::Type idxTy = lowerTy().indexType(); | ||||||
| 2299 | unsigned indexOffset = coor.indicesOffset(); | ||||||
| 2300 | unsigned shapeOffset = coor.shapeOffset(); | ||||||
| 2301 | unsigned shiftOffset = coor.shiftOffset(); | ||||||
| 2302 | unsigned sliceOffset = coor.sliceOffset(); | ||||||
| 2303 | auto sliceOps = coor.getSlice().begin(); | ||||||
| 2304 | mlir::Value one = genConstantIndex(loc, idxTy, rewriter, 1); | ||||||
| 2305 | mlir::Value prevExt = one; | ||||||
| 2306 | mlir::Value offset = genConstantIndex(loc, idxTy, rewriter, 0); | ||||||
| 2307 | const bool isShifted = !coor.getShift().empty(); | ||||||
| 2308 | const bool isSliced = !coor.getSlice().empty(); | ||||||
| 2309 | const bool baseIsBoxed = coor.getMemref().getType().isa<fir::BaseBoxType>(); | ||||||
| 2310 | |||||||
| 2311 | // For each dimension of the array, generate the offset calculation. | ||||||
| 2312 | for (unsigned i = 0; i < rank; ++i, ++indexOffset, ++shapeOffset, | ||||||
| 2313 | ++shiftOffset, sliceOffset += 3, sliceOps += 3) { | ||||||
| 2314 | mlir::Value index = | ||||||
| 2315 | integerCast(loc, rewriter, idxTy, operands[indexOffset]); | ||||||
| 2316 | mlir::Value lb = | ||||||
| 2317 | isShifted ? integerCast(loc, rewriter, idxTy, operands[shiftOffset]) | ||||||
| 2318 | : one; | ||||||
| 2319 | mlir::Value step = one; | ||||||
| 2320 | bool normalSlice = isSliced; | ||||||
| 2321 | // Compute zero based index in dimension i of the element, applying | ||||||
| 2322 | // potential triplets and lower bounds. | ||||||
| 2323 | if (isSliced) { | ||||||
| 2324 | mlir::Value originalUb = *(sliceOps + 1); | ||||||
| 2325 | normalSlice = | ||||||
| 2326 | !mlir::isa_and_nonnull<fir::UndefOp>(originalUb.getDefiningOp()); | ||||||
| 2327 | if (normalSlice) | ||||||
| 2328 | step = integerCast(loc, rewriter, idxTy, operands[sliceOffset + 2]); | ||||||
| 2329 | } | ||||||
| 2330 | auto idx = rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, index, lb); | ||||||
| 2331 | mlir::Value diff = | ||||||
| 2332 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, idx, step); | ||||||
| 2333 | if (normalSlice) { | ||||||
| 2334 | mlir::Value sliceLb = | ||||||
| 2335 | integerCast(loc, rewriter, idxTy, operands[sliceOffset]); | ||||||
| 2336 | auto adj = rewriter.create<mlir::LLVM::SubOp>(loc, idxTy, sliceLb, lb); | ||||||
| 2337 | diff = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, diff, adj); | ||||||
| 2338 | } | ||||||
| 2339 | // Update the offset given the stride and the zero based index `diff` | ||||||
| 2340 | // that was just computed. | ||||||
| 2341 | if (baseIsBoxed) { | ||||||
| 2342 | // Use stride in bytes from the descriptor. | ||||||
| 2343 | mlir::Value stride = getStrideFromBox(loc, coor.getMemref().getType(), | ||||||
| 2344 | operands[0], i, rewriter); | ||||||
| 2345 | auto sc = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, stride); | ||||||
| 2346 | offset = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, offset); | ||||||
| 2347 | } else { | ||||||
| 2348 | // Use stride computed at last iteration. | ||||||
| 2349 | auto sc = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, diff, prevExt); | ||||||
| 2350 | offset = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, offset); | ||||||
| 2351 | // Compute next stride assuming contiguity of the base array | ||||||
| 2352 | // (in element number). | ||||||
| 2353 | auto nextExt = integerCast(loc, rewriter, idxTy, operands[shapeOffset]); | ||||||
| 2354 | prevExt = | ||||||
| 2355 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, prevExt, nextExt); | ||||||
| 2356 | } | ||||||
| 2357 | } | ||||||
| 2358 | |||||||
| 2359 | // Add computed offset to the base address. | ||||||
| 2360 | if (baseIsBoxed) { | ||||||
| 2361 | // Working with byte offsets. The base address is read from the fir.box. | ||||||
| 2362 | // and need to be casted to i8* to do the pointer arithmetic. | ||||||
| 2363 | mlir::Type baseTy = getBaseAddrTypeFromBox(operands[0].getType()); | ||||||
| 2364 | mlir::Value base = getBaseAddrFromBox( | ||||||
| 2365 | loc, baseTy, coor.getMemref().getType(), operands[0], rewriter); | ||||||
| 2366 | mlir::Type voidPtrTy = getVoidPtrType(); | ||||||
| 2367 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, voidPtrTy, base); | ||||||
| 2368 | llvm::SmallVector<mlir::LLVM::GEPArg> args{offset}; | ||||||
| 2369 | auto addr = | ||||||
| 2370 | rewriter.create<mlir::LLVM::GEPOp>(loc, voidPtrTy, base, args); | ||||||
| 2371 | if (coor.getSubcomponent().empty()) { | ||||||
| 2372 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(coor, ty, addr); | ||||||
| 2373 | return mlir::success(); | ||||||
| 2374 | } | ||||||
| 2375 | // Cast the element address from void* to the derived type so that the | ||||||
| 2376 | // derived type members can be addresses via a GEP using the index of | ||||||
| 2377 | // components. | ||||||
| 2378 | mlir::Type elementType = | ||||||
| 2379 | baseTy.cast<mlir::LLVM::LLVMPointerType>().getElementType(); | ||||||
| 2380 | while (auto arrayTy = elementType.dyn_cast<mlir::LLVM::LLVMArrayType>()) | ||||||
| 2381 | elementType = arrayTy.getElementType(); | ||||||
| 2382 | mlir::Type elementPtrType = mlir::LLVM::LLVMPointerType::get(elementType); | ||||||
| 2383 | auto casted = | ||||||
| 2384 | rewriter.create<mlir::LLVM::BitcastOp>(loc, elementPtrType, addr); | ||||||
| 2385 | args.clear(); | ||||||
| 2386 | args.push_back(0); | ||||||
| 2387 | if (!coor.getLenParams().empty()) { | ||||||
| 2388 | // If type parameters are present, then we don't want to use a GEPOp | ||||||
| 2389 | // as below, as the LLVM struct type cannot be statically defined. | ||||||
| 2390 | TODO(loc, "derived type with type parameters")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2390" ": not yet implemented: ") + llvm::Twine("derived type with type parameters" ), false); } while (false); | ||||||
| 2391 | } | ||||||
| 2392 | // TODO: array offset subcomponents must be converted to LLVM's | ||||||
| 2393 | // row-major layout here. | ||||||
| 2394 | for (auto i = coor.subcomponentOffset(); i != coor.indicesOffset(); ++i) | ||||||
| 2395 | args.push_back(operands[i]); | ||||||
| 2396 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(coor, ty, casted, args); | ||||||
| 2397 | return mlir::success(); | ||||||
| 2398 | } | ||||||
| 2399 | |||||||
| 2400 | // The array was not boxed, so it must be contiguous. offset is therefore an | ||||||
| 2401 | // element offset and the base type is kept in the GEP unless the element | ||||||
| 2402 | // type size is itself dynamic. | ||||||
| 2403 | mlir::Value base; | ||||||
| 2404 | if (coor.getSubcomponent().empty()) { | ||||||
| 2405 | // No subcomponent. | ||||||
| 2406 | if (!coor.getLenParams().empty()) { | ||||||
| 2407 | // Type parameters. Adjust element size explicitly. | ||||||
| 2408 | auto eleTy = fir::dyn_cast_ptrEleTy(coor.getType()); | ||||||
| 2409 | assert(eleTy && "result must be a reference-like type")(static_cast <bool> (eleTy && "result must be a reference-like type" ) ? void (0) : __assert_fail ("eleTy && \"result must be a reference-like type\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2409, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2410 | if (fir::characterWithDynamicLen(eleTy)) { | ||||||
| 2411 | assert(coor.getLenParams().size() == 1)(static_cast <bool> (coor.getLenParams().size() == 1) ? void (0) : __assert_fail ("coor.getLenParams().size() == 1", "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2411, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2412 | auto length = integerCast(loc, rewriter, idxTy, | ||||||
| 2413 | operands[coor.lenParamsOffset()]); | ||||||
| 2414 | offset = | ||||||
| 2415 | rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, offset, length); | ||||||
| 2416 | } else { | ||||||
| 2417 | TODO(loc, "compute size of derived type with type parameters")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2417" ": not yet implemented: ") + llvm::Twine("compute size of derived type with type parameters" ), false); } while (false); | ||||||
| 2418 | } | ||||||
| 2419 | } | ||||||
| 2420 | // Cast the base address to a pointer to T. | ||||||
| 2421 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, ty, operands[0]); | ||||||
| 2422 | } else { | ||||||
| 2423 | // Operand #0 must have a pointer type. For subcomponent slicing, we | ||||||
| 2424 | // want to cast away the array type and have a plain struct type. | ||||||
| 2425 | mlir::Type ty0 = operands[0].getType(); | ||||||
| 2426 | auto ptrTy = ty0.dyn_cast<mlir::LLVM::LLVMPointerType>(); | ||||||
| 2427 | assert(ptrTy && "expected pointer type")(static_cast <bool> (ptrTy && "expected pointer type" ) ? void (0) : __assert_fail ("ptrTy && \"expected pointer type\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2427, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2428 | mlir::Type eleTy = ptrTy.getElementType(); | ||||||
| 2429 | while (auto arrTy = eleTy.dyn_cast<mlir::LLVM::LLVMArrayType>()) | ||||||
| 2430 | eleTy = arrTy.getElementType(); | ||||||
| 2431 | auto newTy = mlir::LLVM::LLVMPointerType::get(eleTy); | ||||||
| 2432 | base = rewriter.create<mlir::LLVM::BitcastOp>(loc, newTy, operands[0]); | ||||||
| 2433 | } | ||||||
| 2434 | llvm::SmallVector<mlir::LLVM::GEPArg> args = {offset}; | ||||||
| 2435 | for (auto i = coor.subcomponentOffset(); i != coor.indicesOffset(); ++i) | ||||||
| 2436 | args.push_back(operands[i]); | ||||||
| 2437 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(coor, ty, base, args); | ||||||
| 2438 | return mlir::success(); | ||||||
| 2439 | } | ||||||
| 2440 | }; | ||||||
| 2441 | } // namespace | ||||||
| 2442 | |||||||
| 2443 | /// Convert to (memory) reference to a reference to a subobject. | ||||||
| 2444 | /// The coordinate_of op is a Swiss army knife operation that can be used on | ||||||
| 2445 | /// (memory) references to records, arrays, complex, etc. as well as boxes. | ||||||
| 2446 | /// With unboxed arrays, there is the restriction that the array have a static | ||||||
| 2447 | /// shape in all but the last column. | ||||||
| 2448 | struct CoordinateOpConversion | ||||||
| 2449 | : public FIROpAndTypeConversion<fir::CoordinateOp> { | ||||||
| 2450 | using FIROpAndTypeConversion::FIROpAndTypeConversion; | ||||||
| 2451 | |||||||
| 2452 | mlir::LogicalResult | ||||||
| 2453 | doRewrite(fir::CoordinateOp coor, mlir::Type ty, OpAdaptor adaptor, | ||||||
| 2454 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2455 | mlir::ValueRange operands = adaptor.getOperands(); | ||||||
| 2456 | |||||||
| 2457 | mlir::Location loc = coor.getLoc(); | ||||||
| 2458 | mlir::Value base = operands[0]; | ||||||
| 2459 | mlir::Type baseObjectTy = coor.getBaseType(); | ||||||
| 2460 | mlir::Type objectTy = fir::dyn_cast_ptrOrBoxEleTy(baseObjectTy); | ||||||
| 2461 | assert(objectTy && "fir.coordinate_of expects a reference type")(static_cast <bool> (objectTy && "fir.coordinate_of expects a reference type" ) ? void (0) : __assert_fail ("objectTy && \"fir.coordinate_of expects a reference type\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2461, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2462 | |||||||
| 2463 | // Complex type - basically, extract the real or imaginary part | ||||||
| 2464 | if (fir::isa_complex(objectTy)) { | ||||||
| 2465 | mlir::Value gep = genGEP(loc, ty, rewriter, base, 0, operands[1]); | ||||||
| 2466 | rewriter.replaceOp(coor, gep); | ||||||
| 2467 | return mlir::success(); | ||||||
| 2468 | } | ||||||
| 2469 | |||||||
| 2470 | // Boxed type - get the base pointer from the box | ||||||
| 2471 | if (baseObjectTy.dyn_cast<fir::BaseBoxType>()) | ||||||
| 2472 | return doRewriteBox(coor, ty, operands, loc, rewriter); | ||||||
| 2473 | |||||||
| 2474 | // Reference, pointer or a heap type | ||||||
| 2475 | if (baseObjectTy.isa<fir::ReferenceType, fir::PointerType, fir::HeapType>()) | ||||||
| 2476 | return doRewriteRefOrPtr(coor, ty, operands, loc, rewriter); | ||||||
| 2477 | |||||||
| 2478 | return rewriter.notifyMatchFailure( | ||||||
| 2479 | coor, "fir.coordinate_of base operand has unsupported type"); | ||||||
| 2480 | } | ||||||
| 2481 | |||||||
| 2482 | static unsigned getFieldNumber(fir::RecordType ty, mlir::Value op) { | ||||||
| 2483 | return fir::hasDynamicSize(ty) | ||||||
| 2484 | ? op.getDefiningOp() | ||||||
| 2485 | ->getAttrOfType<mlir::IntegerAttr>("field") | ||||||
| 2486 | .getInt() | ||||||
| 2487 | : getConstantIntValue(op); | ||||||
| 2488 | } | ||||||
| 2489 | |||||||
| 2490 | static bool hasSubDimensions(mlir::Type type) { | ||||||
| 2491 | return type.isa<fir::SequenceType, fir::RecordType, mlir::TupleType>(); | ||||||
| 2492 | } | ||||||
| 2493 | |||||||
| 2494 | /// Check whether this form of `!fir.coordinate_of` is supported. These | ||||||
| 2495 | /// additional checks are required, because we are not yet able to convert | ||||||
| 2496 | /// all valid forms of `!fir.coordinate_of`. | ||||||
| 2497 | /// TODO: Either implement the unsupported cases or extend the verifier | ||||||
| 2498 | /// in FIROps.cpp instead. | ||||||
| 2499 | static bool supportedCoordinate(mlir::Type type, mlir::ValueRange coors) { | ||||||
| 2500 | const std::size_t numOfCoors = coors.size(); | ||||||
| 2501 | std::size_t i = 0; | ||||||
| 2502 | bool subEle = false; | ||||||
| 2503 | bool ptrEle = false; | ||||||
| 2504 | for (; i < numOfCoors; ++i) { | ||||||
| 2505 | mlir::Value nxtOpnd = coors[i]; | ||||||
| 2506 | if (auto arrTy = type.dyn_cast<fir::SequenceType>()) { | ||||||
| 2507 | subEle = true; | ||||||
| 2508 | i += arrTy.getDimension() - 1; | ||||||
| 2509 | type = arrTy.getEleTy(); | ||||||
| 2510 | } else if (auto recTy = type.dyn_cast<fir::RecordType>()) { | ||||||
| 2511 | subEle = true; | ||||||
| 2512 | type = recTy.getType(getFieldNumber(recTy, nxtOpnd)); | ||||||
| 2513 | } else if (auto tupTy = type.dyn_cast<mlir::TupleType>()) { | ||||||
| 2514 | subEle = true; | ||||||
| 2515 | type = tupTy.getType(getConstantIntValue(nxtOpnd)); | ||||||
| 2516 | } else { | ||||||
| 2517 | ptrEle = true; | ||||||
| 2518 | } | ||||||
| 2519 | } | ||||||
| 2520 | if (ptrEle) | ||||||
| 2521 | return (!subEle) && (numOfCoors == 1); | ||||||
| 2522 | return subEle && (i >= numOfCoors); | ||||||
| 2523 | } | ||||||
| 2524 | |||||||
| 2525 | /// Walk the abstract memory layout and determine if the path traverses any | ||||||
| 2526 | /// array types with unknown shape. Return true iff all the array types have a | ||||||
| 2527 | /// constant shape along the path. | ||||||
| 2528 | static bool arraysHaveKnownShape(mlir::Type type, mlir::ValueRange coors) { | ||||||
| 2529 | for (std::size_t i = 0, sz = coors.size(); i < sz; ++i) { | ||||||
| 2530 | mlir::Value nxtOpnd = coors[i]; | ||||||
| 2531 | if (auto arrTy = type.dyn_cast<fir::SequenceType>()) { | ||||||
| 2532 | if (fir::sequenceWithNonConstantShape(arrTy)) | ||||||
| 2533 | return false; | ||||||
| 2534 | i += arrTy.getDimension() - 1; | ||||||
| 2535 | type = arrTy.getEleTy(); | ||||||
| 2536 | } else if (auto strTy = type.dyn_cast<fir::RecordType>()) { | ||||||
| 2537 | type = strTy.getType(getFieldNumber(strTy, nxtOpnd)); | ||||||
| 2538 | } else if (auto strTy = type.dyn_cast<mlir::TupleType>()) { | ||||||
| 2539 | type = strTy.getType(getConstantIntValue(nxtOpnd)); | ||||||
| 2540 | } else { | ||||||
| 2541 | return true; | ||||||
| 2542 | } | ||||||
| 2543 | } | ||||||
| 2544 | return true; | ||||||
| 2545 | } | ||||||
| 2546 | |||||||
| 2547 | private: | ||||||
| 2548 | mlir::LogicalResult | ||||||
| 2549 | doRewriteBox(fir::CoordinateOp coor, mlir::Type ty, mlir::ValueRange operands, | ||||||
| 2550 | mlir::Location loc, | ||||||
| 2551 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 2552 | mlir::Type boxObjTy = coor.getBaseType(); | ||||||
| 2553 | assert(boxObjTy.dyn_cast<fir::BaseBoxType>() && "This is not a `fir.box`")(static_cast <bool> (boxObjTy.dyn_cast<fir::BaseBoxType >() && "This is not a `fir.box`") ? void (0) : __assert_fail ("boxObjTy.dyn_cast<fir::BaseBoxType>() && \"This is not a `fir.box`\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2553, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2554 | |||||||
| 2555 | mlir::Value boxBaseAddr = operands[0]; | ||||||
| 2556 | |||||||
| 2557 | // 1. SPECIAL CASE (uses `fir.len_param_index`): | ||||||
| 2558 | // %box = ... : !fir.box<!fir.type<derived{len1:i32}>> | ||||||
| 2559 | // %lenp = fir.len_param_index len1, !fir.type<derived{len1:i32}> | ||||||
| 2560 | // %addr = coordinate_of %box, %lenp | ||||||
| 2561 | if (coor.getNumOperands() == 2) { | ||||||
| 2562 | mlir::Operation *coordinateDef = | ||||||
| 2563 | (*coor.getCoor().begin()).getDefiningOp(); | ||||||
| 2564 | if (mlir::isa_and_nonnull<fir::LenParamIndexOp>(coordinateDef)) | ||||||
| 2565 | TODO(loc,do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2566" ": not yet implemented: ") + llvm::Twine("fir.coordinate_of - fir.len_param_index is not supported yet" ), false); } while (false) | ||||||
| 2566 | "fir.coordinate_of - fir.len_param_index is not supported yet")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2566" ": not yet implemented: ") + llvm::Twine("fir.coordinate_of - fir.len_param_index is not supported yet" ), false); } while (false); | ||||||
| 2567 | } | ||||||
| 2568 | |||||||
| 2569 | // 2. GENERAL CASE: | ||||||
| 2570 | // 2.1. (`fir.array`) | ||||||
| 2571 | // %box = ... : !fix.box<!fir.array<?xU>> | ||||||
| 2572 | // %idx = ... : index | ||||||
| 2573 | // %resultAddr = coordinate_of %box, %idx : !fir.ref<U> | ||||||
| 2574 | // 2.2 (`fir.derived`) | ||||||
| 2575 | // %box = ... : !fix.box<!fir.type<derived_type{field_1:i32}>> | ||||||
| 2576 | // %idx = ... : i32 | ||||||
| 2577 | // %resultAddr = coordinate_of %box, %idx : !fir.ref<i32> | ||||||
| 2578 | // 2.3 (`fir.derived` inside `fir.array`) | ||||||
| 2579 | // %box = ... : !fir.box<!fir.array<10 x !fir.type<derived_1{field_1:f32, | ||||||
| 2580 | // field_2:f32}>>> %idx1 = ... : index %idx2 = ... : i32 %resultAddr = | ||||||
| 2581 | // coordinate_of %box, %idx1, %idx2 : !fir.ref<f32> | ||||||
| 2582 | // 2.4. TODO: Either document or disable any other case that the following | ||||||
| 2583 | // implementation might convert. | ||||||
| 2584 | mlir::Value resultAddr = | ||||||
| 2585 | getBaseAddrFromBox(loc, getBaseAddrTypeFromBox(boxBaseAddr.getType()), | ||||||
| 2586 | boxObjTy, boxBaseAddr, rewriter); | ||||||
| 2587 | // Component Type | ||||||
| 2588 | auto cpnTy = fir::dyn_cast_ptrOrBoxEleTy(boxObjTy); | ||||||
| 2589 | mlir::Type voidPtrTy = ::getVoidPtrType(coor.getContext()); | ||||||
| 2590 | |||||||
| 2591 | for (unsigned i = 1, last = operands.size(); i < last; ++i) { | ||||||
| 2592 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { | ||||||
| 2593 | if (i != 1) | ||||||
| 2594 | TODO(loc, "fir.array nested inside other array and/or derived type")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2594" ": not yet implemented: ") + llvm::Twine("fir.array nested inside other array and/or derived type" ), false); } while (false); | ||||||
| 2595 | // Applies byte strides from the box. Ignore lower bound from box | ||||||
| 2596 | // since fir.coordinate_of indexes are zero based. Lowering takes care | ||||||
| 2597 | // of lower bound aspects. This both accounts for dynamically sized | ||||||
| 2598 | // types and non contiguous arrays. | ||||||
| 2599 | auto idxTy = lowerTy().indexType(); | ||||||
| 2600 | mlir::Value off = genConstantIndex(loc, idxTy, rewriter, 0); | ||||||
| 2601 | for (unsigned index = i, lastIndex = i + arrTy.getDimension(); | ||||||
| 2602 | index < lastIndex; ++index) { | ||||||
| 2603 | mlir::Value stride = | ||||||
| 2604 | getStrideFromBox(loc, boxObjTy, operands[0], index - i, rewriter); | ||||||
| 2605 | auto sc = rewriter.create<mlir::LLVM::MulOp>(loc, idxTy, | ||||||
| 2606 | operands[index], stride); | ||||||
| 2607 | off = rewriter.create<mlir::LLVM::AddOp>(loc, idxTy, sc, off); | ||||||
| 2608 | } | ||||||
| 2609 | auto voidPtrBase = | ||||||
| 2610 | rewriter.create<mlir::LLVM::BitcastOp>(loc, voidPtrTy, resultAddr); | ||||||
| 2611 | resultAddr = rewriter.create<mlir::LLVM::GEPOp>( | ||||||
| 2612 | loc, voidPtrTy, voidPtrBase, | ||||||
| 2613 | llvm::ArrayRef<mlir::LLVM::GEPArg>{off}); | ||||||
| 2614 | i += arrTy.getDimension() - 1; | ||||||
| 2615 | cpnTy = arrTy.getEleTy(); | ||||||
| 2616 | } else if (auto recTy = cpnTy.dyn_cast<fir::RecordType>()) { | ||||||
| 2617 | auto recRefTy = | ||||||
| 2618 | mlir::LLVM::LLVMPointerType::get(lowerTy().convertType(recTy)); | ||||||
| 2619 | mlir::Value nxtOpnd = operands[i]; | ||||||
| 2620 | auto memObj = | ||||||
| 2621 | rewriter.create<mlir::LLVM::BitcastOp>(loc, recRefTy, resultAddr); | ||||||
| 2622 | cpnTy = recTy.getType(getFieldNumber(recTy, nxtOpnd)); | ||||||
| 2623 | auto llvmCurrentObjTy = lowerTy().convertType(cpnTy); | ||||||
| 2624 | auto gep = rewriter.create<mlir::LLVM::GEPOp>( | ||||||
| 2625 | loc, mlir::LLVM::LLVMPointerType::get(llvmCurrentObjTy), memObj, | ||||||
| 2626 | llvm::ArrayRef<mlir::LLVM::GEPArg>{0, nxtOpnd}); | ||||||
| 2627 | resultAddr = | ||||||
| 2628 | rewriter.create<mlir::LLVM::BitcastOp>(loc, voidPtrTy, gep); | ||||||
| 2629 | } else { | ||||||
| 2630 | fir::emitFatalError(loc, "unexpected type in coordinate_of"); | ||||||
| 2631 | } | ||||||
| 2632 | } | ||||||
| 2633 | |||||||
| 2634 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(coor, ty, resultAddr); | ||||||
| 2635 | return mlir::success(); | ||||||
| 2636 | } | ||||||
| 2637 | |||||||
| 2638 | mlir::LogicalResult | ||||||
| 2639 | doRewriteRefOrPtr(fir::CoordinateOp coor, mlir::Type ty, | ||||||
| 2640 | mlir::ValueRange operands, mlir::Location loc, | ||||||
| 2641 | mlir::ConversionPatternRewriter &rewriter) const { | ||||||
| 2642 | mlir::Type baseObjectTy = coor.getBaseType(); | ||||||
| 2643 | |||||||
| 2644 | // Component Type | ||||||
| 2645 | mlir::Type cpnTy = fir::dyn_cast_ptrOrBoxEleTy(baseObjectTy); | ||||||
| 2646 | bool hasSubdimension = hasSubDimensions(cpnTy); | ||||||
| 2647 | bool columnIsDeferred = !hasSubdimension; | ||||||
| 2648 | |||||||
| 2649 | if (!supportedCoordinate(cpnTy, operands.drop_front(1))) | ||||||
| 2650 | TODO(loc, "unsupported combination of coordinate operands")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2650" ": not yet implemented: ") + llvm::Twine("unsupported combination of coordinate operands" ), false); } while (false); | ||||||
| 2651 | |||||||
| 2652 | const bool hasKnownShape = | ||||||
| 2653 | arraysHaveKnownShape(cpnTy, operands.drop_front(1)); | ||||||
| 2654 | |||||||
| 2655 | // If only the column is `?`, then we can simply place the column value in | ||||||
| 2656 | // the 0-th GEP position. | ||||||
| 2657 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { | ||||||
| 2658 | if (!hasKnownShape) { | ||||||
| 2659 | const unsigned sz = arrTy.getDimension(); | ||||||
| 2660 | if (arraysHaveKnownShape(arrTy.getEleTy(), | ||||||
| 2661 | operands.drop_front(1 + sz))) { | ||||||
| 2662 | fir::SequenceType::ShapeRef shape = arrTy.getShape(); | ||||||
| 2663 | bool allConst = true; | ||||||
| 2664 | for (unsigned i = 0; i < sz - 1; ++i) { | ||||||
| 2665 | if (shape[i] < 0) { | ||||||
| 2666 | allConst = false; | ||||||
| 2667 | break; | ||||||
| 2668 | } | ||||||
| 2669 | } | ||||||
| 2670 | if (allConst) | ||||||
| 2671 | columnIsDeferred = true; | ||||||
| 2672 | } | ||||||
| 2673 | } | ||||||
| 2674 | } | ||||||
| 2675 | |||||||
| 2676 | if (fir::hasDynamicSize(fir::unwrapSequenceType(cpnTy))) | ||||||
| 2677 | return mlir::emitError( | ||||||
| 2678 | loc, "fir.coordinate_of with a dynamic element size is unsupported"); | ||||||
| 2679 | |||||||
| 2680 | if (hasKnownShape || columnIsDeferred) { | ||||||
| 2681 | llvm::SmallVector<mlir::LLVM::GEPArg> offs; | ||||||
| 2682 | if (hasKnownShape && hasSubdimension) { | ||||||
| 2683 | offs.push_back(0); | ||||||
| 2684 | } | ||||||
| 2685 | std::optional<int> dims; | ||||||
| 2686 | llvm::SmallVector<mlir::Value> arrIdx; | ||||||
| 2687 | for (std::size_t i = 1, sz = operands.size(); i < sz; ++i) { | ||||||
| 2688 | mlir::Value nxtOpnd = operands[i]; | ||||||
| 2689 | |||||||
| 2690 | if (!cpnTy) | ||||||
| 2691 | return mlir::emitError(loc, "invalid coordinate/check failed"); | ||||||
| 2692 | |||||||
| 2693 | // check if the i-th coordinate relates to an array | ||||||
| 2694 | if (dims) { | ||||||
| 2695 | arrIdx.push_back(nxtOpnd); | ||||||
| 2696 | int dimsLeft = *dims; | ||||||
| 2697 | if (dimsLeft > 1) { | ||||||
| 2698 | dims = dimsLeft - 1; | ||||||
| 2699 | continue; | ||||||
| 2700 | } | ||||||
| 2701 | cpnTy = cpnTy.cast<fir::SequenceType>().getEleTy(); | ||||||
| 2702 | // append array range in reverse (FIR arrays are column-major) | ||||||
| 2703 | offs.append(arrIdx.rbegin(), arrIdx.rend()); | ||||||
| 2704 | arrIdx.clear(); | ||||||
| 2705 | dims.reset(); | ||||||
| 2706 | continue; | ||||||
| 2707 | } | ||||||
| 2708 | if (auto arrTy = cpnTy.dyn_cast<fir::SequenceType>()) { | ||||||
| 2709 | int d = arrTy.getDimension() - 1; | ||||||
| 2710 | if (d > 0) { | ||||||
| 2711 | dims = d; | ||||||
| 2712 | arrIdx.push_back(nxtOpnd); | ||||||
| 2713 | continue; | ||||||
| 2714 | } | ||||||
| 2715 | cpnTy = cpnTy.cast<fir::SequenceType>().getEleTy(); | ||||||
| 2716 | offs.push_back(nxtOpnd); | ||||||
| 2717 | continue; | ||||||
| 2718 | } | ||||||
| 2719 | |||||||
| 2720 | // check if the i-th coordinate relates to a field | ||||||
| 2721 | if (auto recTy = cpnTy.dyn_cast<fir::RecordType>()) | ||||||
| 2722 | cpnTy = recTy.getType(getFieldNumber(recTy, nxtOpnd)); | ||||||
| 2723 | else if (auto tupTy = cpnTy.dyn_cast<mlir::TupleType>()) | ||||||
| 2724 | cpnTy = tupTy.getType(getConstantIntValue(nxtOpnd)); | ||||||
| 2725 | else | ||||||
| 2726 | cpnTy = nullptr; | ||||||
| 2727 | |||||||
| 2728 | offs.push_back(nxtOpnd); | ||||||
| 2729 | } | ||||||
| 2730 | if (dims) | ||||||
| 2731 | offs.append(arrIdx.rbegin(), arrIdx.rend()); | ||||||
| 2732 | mlir::Value base = operands[0]; | ||||||
| 2733 | mlir::Value retval = genGEP(loc, ty, rewriter, base, offs); | ||||||
| 2734 | rewriter.replaceOp(coor, retval); | ||||||
| 2735 | return mlir::success(); | ||||||
| 2736 | } | ||||||
| 2737 | |||||||
| 2738 | return mlir::emitError( | ||||||
| 2739 | loc, "fir.coordinate_of base operand has unsupported type"); | ||||||
| 2740 | } | ||||||
| 2741 | }; | ||||||
| 2742 | |||||||
| 2743 | /// Convert `fir.field_index`. The conversion depends on whether the size of | ||||||
| 2744 | /// the record is static or dynamic. | ||||||
| 2745 | struct FieldIndexOpConversion : public FIROpConversion<fir::FieldIndexOp> { | ||||||
| 2746 | using FIROpConversion::FIROpConversion; | ||||||
| 2747 | |||||||
| 2748 | // NB: most field references should be resolved by this point | ||||||
| 2749 | mlir::LogicalResult | ||||||
| 2750 | matchAndRewrite(fir::FieldIndexOp field, OpAdaptor adaptor, | ||||||
| 2751 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2752 | auto recTy = field.getOnType().cast<fir::RecordType>(); | ||||||
| 2753 | unsigned index = recTy.getFieldIndex(field.getFieldId()); | ||||||
| 2754 | |||||||
| 2755 | if (!fir::hasDynamicSize(recTy)) { | ||||||
| 2756 | // Derived type has compile-time constant layout. Return index of the | ||||||
| 2757 | // component type in the parent type (to be used in GEP). | ||||||
| 2758 | rewriter.replaceOp(field, mlir::ValueRange{genConstantOffset( | ||||||
| 2759 | field.getLoc(), rewriter, index)}); | ||||||
| 2760 | return mlir::success(); | ||||||
| 2761 | } | ||||||
| 2762 | |||||||
| 2763 | // Derived type has compile-time constant layout. Call the compiler | ||||||
| 2764 | // generated function to determine the byte offset of the field at runtime. | ||||||
| 2765 | // This returns a non-constant. | ||||||
| 2766 | mlir::FlatSymbolRefAttr symAttr = mlir::SymbolRefAttr::get( | ||||||
| 2767 | field.getContext(), getOffsetMethodName(recTy, field.getFieldId())); | ||||||
| 2768 | mlir::NamedAttribute callAttr = rewriter.getNamedAttr("callee", symAttr); | ||||||
| 2769 | mlir::NamedAttribute fieldAttr = rewriter.getNamedAttr( | ||||||
| 2770 | "field", mlir::IntegerAttr::get(lowerTy().indexType(), index)); | ||||||
| 2771 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( | ||||||
| 2772 | field, lowerTy().offsetType(), adaptor.getOperands(), | ||||||
| 2773 | llvm::ArrayRef<mlir::NamedAttribute>{callAttr, fieldAttr}); | ||||||
| 2774 | return mlir::success(); | ||||||
| 2775 | } | ||||||
| 2776 | |||||||
| 2777 | // Re-Construct the name of the compiler generated method that calculates the | ||||||
| 2778 | // offset | ||||||
| 2779 | inline static std::string getOffsetMethodName(fir::RecordType recTy, | ||||||
| 2780 | llvm::StringRef field) { | ||||||
| 2781 | return recTy.getName().str() + "P." + field.str() + ".offset"; | ||||||
| 2782 | } | ||||||
| 2783 | }; | ||||||
| 2784 | |||||||
| 2785 | /// Convert `fir.end` | ||||||
| 2786 | struct FirEndOpConversion : public FIROpConversion<fir::FirEndOp> { | ||||||
| 2787 | using FIROpConversion::FIROpConversion; | ||||||
| 2788 | |||||||
| 2789 | mlir::LogicalResult | ||||||
| 2790 | matchAndRewrite(fir::FirEndOp firEnd, OpAdaptor, | ||||||
| 2791 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2792 | TODO(firEnd.getLoc(), "fir.end codegen")do { fir::emitFatalError(firEnd.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2792" ": not yet implemented: ") + llvm::Twine("fir.end codegen" ), false); } while (false); | ||||||
| 2793 | return mlir::failure(); | ||||||
| 2794 | } | ||||||
| 2795 | }; | ||||||
| 2796 | |||||||
| 2797 | /// Lower `fir.type_desc` to a global addr. | ||||||
| 2798 | struct TypeDescOpConversion : public FIROpConversion<fir::TypeDescOp> { | ||||||
| 2799 | using FIROpConversion::FIROpConversion; | ||||||
| 2800 | |||||||
| 2801 | mlir::LogicalResult | ||||||
| 2802 | matchAndRewrite(fir::TypeDescOp typeDescOp, OpAdaptor adaptor, | ||||||
| 2803 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2804 | mlir::Type inTy = typeDescOp.getInType(); | ||||||
| 2805 | assert(inTy.isa<fir::RecordType>() && "expecting fir.type")(static_cast <bool> (inTy.isa<fir::RecordType>() && "expecting fir.type") ? void (0) : __assert_fail ("inTy.isa<fir::RecordType>() && \"expecting fir.type\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 2805, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2806 | auto recordType = inTy.dyn_cast<fir::RecordType>(); | ||||||
| 2807 | auto module = typeDescOp.getOperation()->getParentOfType<mlir::ModuleOp>(); | ||||||
| 2808 | std::string typeDescName = | ||||||
| 2809 | fir::NameUniquer::getTypeDescriptorName(recordType.getName()); | ||||||
| 2810 | if (auto global = module.lookupSymbol<mlir::LLVM::GlobalOp>(typeDescName)) { | ||||||
| 2811 | auto ty = mlir::LLVM::LLVMPointerType::get( | ||||||
| 2812 | this->lowerTy().convertType(global.getType())); | ||||||
| 2813 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>(typeDescOp, ty, | ||||||
| 2814 | global.getSymName()); | ||||||
| 2815 | return mlir::success(); | ||||||
| 2816 | } else if (auto global = module.lookupSymbol<fir::GlobalOp>(typeDescName)) { | ||||||
| 2817 | auto ty = mlir::LLVM::LLVMPointerType::get( | ||||||
| 2818 | this->lowerTy().convertType(global.getType())); | ||||||
| 2819 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddressOfOp>(typeDescOp, ty, | ||||||
| 2820 | global.getSymName()); | ||||||
| 2821 | return mlir::success(); | ||||||
| 2822 | } | ||||||
| 2823 | return mlir::failure(); | ||||||
| 2824 | } | ||||||
| 2825 | }; | ||||||
| 2826 | |||||||
| 2827 | /// Lower `fir.has_value` operation to `llvm.return` operation. | ||||||
| 2828 | struct HasValueOpConversion : public FIROpConversion<fir::HasValueOp> { | ||||||
| 2829 | using FIROpConversion::FIROpConversion; | ||||||
| 2830 | |||||||
| 2831 | mlir::LogicalResult | ||||||
| 2832 | matchAndRewrite(fir::HasValueOp op, OpAdaptor adaptor, | ||||||
| 2833 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2834 | rewriter.replaceOpWithNewOp<mlir::LLVM::ReturnOp>(op, | ||||||
| 2835 | adaptor.getOperands()); | ||||||
| 2836 | return mlir::success(); | ||||||
| 2837 | } | ||||||
| 2838 | }; | ||||||
| 2839 | |||||||
| 2840 | /// Lower `fir.global` operation to `llvm.global` operation. | ||||||
| 2841 | /// `fir.insert_on_range` operations are replaced with constant dense attribute | ||||||
| 2842 | /// if they are applied on the full range. | ||||||
| 2843 | struct GlobalOpConversion : public FIROpConversion<fir::GlobalOp> { | ||||||
| 2844 | using FIROpConversion::FIROpConversion; | ||||||
| 2845 | |||||||
| 2846 | mlir::LogicalResult | ||||||
| 2847 | matchAndRewrite(fir::GlobalOp global, OpAdaptor adaptor, | ||||||
| 2848 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2849 | auto tyAttr = convertType(global.getType()); | ||||||
| 2850 | if (global.getType().isa<fir::BaseBoxType>()) | ||||||
| 2851 | tyAttr = tyAttr.cast<mlir::LLVM::LLVMPointerType>().getElementType(); | ||||||
| 2852 | auto loc = global.getLoc(); | ||||||
| 2853 | mlir::Attribute initAttr = global.getInitVal().value_or(mlir::Attribute()); | ||||||
| 2854 | auto linkage = convertLinkage(global.getLinkName()); | ||||||
| 2855 | auto isConst = global.getConstant().has_value(); | ||||||
| 2856 | auto g = rewriter.create<mlir::LLVM::GlobalOp>( | ||||||
| 2857 | loc, tyAttr, isConst, linkage, global.getSymName(), initAttr); | ||||||
| 2858 | |||||||
| 2859 | // Apply all non-Fir::GlobalOp attributes to the LLVM::GlobalOp, preserving | ||||||
| 2860 | // them; whilst taking care not to apply attributes that are lowered in | ||||||
| 2861 | // other ways. | ||||||
| 2862 | llvm::SmallDenseSet<llvm::StringRef> elidedAttrsSet( | ||||||
| 2863 | global.getAttributeNames().begin(), global.getAttributeNames().end()); | ||||||
| 2864 | for (auto &attr : global->getAttrs()) | ||||||
| 2865 | if (!elidedAttrsSet.contains(attr.getName().strref())) | ||||||
| 2866 | g->setAttr(attr.getName(), attr.getValue()); | ||||||
| 2867 | |||||||
| 2868 | auto &gr = g.getInitializerRegion(); | ||||||
| 2869 | rewriter.inlineRegionBefore(global.getRegion(), gr, gr.end()); | ||||||
| 2870 | if (!gr.empty()) { | ||||||
| 2871 | // Replace insert_on_range with a constant dense attribute if the | ||||||
| 2872 | // initialization is on the full range. | ||||||
| 2873 | auto insertOnRangeOps = gr.front().getOps<fir::InsertOnRangeOp>(); | ||||||
| 2874 | for (auto insertOp : insertOnRangeOps) { | ||||||
| 2875 | if (isFullRange(insertOp.getCoor(), insertOp.getType())) { | ||||||
| 2876 | auto seqTyAttr = convertType(insertOp.getType()); | ||||||
| 2877 | auto *op = insertOp.getVal().getDefiningOp(); | ||||||
| 2878 | auto constant = mlir::dyn_cast<mlir::arith::ConstantOp>(op); | ||||||
| 2879 | if (!constant) { | ||||||
| 2880 | auto convertOp = mlir::dyn_cast<fir::ConvertOp>(op); | ||||||
| 2881 | if (!convertOp) | ||||||
| 2882 | continue; | ||||||
| 2883 | constant = mlir::cast<mlir::arith::ConstantOp>( | ||||||
| 2884 | convertOp.getValue().getDefiningOp()); | ||||||
| 2885 | } | ||||||
| 2886 | mlir::Type vecType = mlir::VectorType::get( | ||||||
| 2887 | insertOp.getType().getShape(), constant.getType()); | ||||||
| 2888 | auto denseAttr = mlir::DenseElementsAttr::get( | ||||||
| 2889 | vecType.cast<mlir::ShapedType>(), constant.getValue()); | ||||||
| 2890 | rewriter.setInsertionPointAfter(insertOp); | ||||||
| 2891 | rewriter.replaceOpWithNewOp<mlir::arith::ConstantOp>( | ||||||
| 2892 | insertOp, seqTyAttr, denseAttr); | ||||||
| 2893 | } | ||||||
| 2894 | } | ||||||
| 2895 | } | ||||||
| 2896 | rewriter.eraseOp(global); | ||||||
| 2897 | return mlir::success(); | ||||||
| 2898 | } | ||||||
| 2899 | |||||||
| 2900 | bool isFullRange(mlir::DenseIntElementsAttr indexes, | ||||||
| 2901 | fir::SequenceType seqTy) const { | ||||||
| 2902 | auto extents = seqTy.getShape(); | ||||||
| 2903 | if (indexes.size() / 2 != static_cast<int64_t>(extents.size())) | ||||||
| 2904 | return false; | ||||||
| 2905 | auto cur_index = indexes.value_begin<int64_t>(); | ||||||
| 2906 | for (unsigned i = 0; i < indexes.size(); i += 2) { | ||||||
| 2907 | if (*(cur_index++) != 0) | ||||||
| 2908 | return false; | ||||||
| 2909 | if (*(cur_index++) != extents[i / 2] - 1) | ||||||
| 2910 | return false; | ||||||
| 2911 | } | ||||||
| 2912 | return true; | ||||||
| 2913 | } | ||||||
| 2914 | |||||||
| 2915 | // TODO: String comparaison should be avoided. Replace linkName with an | ||||||
| 2916 | // enumeration. | ||||||
| 2917 | mlir::LLVM::Linkage | ||||||
| 2918 | convertLinkage(std::optional<llvm::StringRef> optLinkage) const { | ||||||
| 2919 | if (optLinkage) { | ||||||
| 2920 | auto name = *optLinkage; | ||||||
| 2921 | if (name == "internal") | ||||||
| 2922 | return mlir::LLVM::Linkage::Internal; | ||||||
| 2923 | if (name == "linkonce") | ||||||
| 2924 | return mlir::LLVM::Linkage::Linkonce; | ||||||
| 2925 | if (name == "linkonce_odr") | ||||||
| 2926 | return mlir::LLVM::Linkage::LinkonceODR; | ||||||
| 2927 | if (name == "common") | ||||||
| 2928 | return mlir::LLVM::Linkage::Common; | ||||||
| 2929 | if (name == "weak") | ||||||
| 2930 | return mlir::LLVM::Linkage::Weak; | ||||||
| 2931 | } | ||||||
| 2932 | return mlir::LLVM::Linkage::External; | ||||||
| 2933 | } | ||||||
| 2934 | }; | ||||||
| 2935 | |||||||
| 2936 | /// `fir.load` --> `llvm.load` | ||||||
| 2937 | struct LoadOpConversion : public FIROpConversion<fir::LoadOp> { | ||||||
| 2938 | using FIROpConversion::FIROpConversion; | ||||||
| 2939 | |||||||
| 2940 | mlir::LogicalResult | ||||||
| 2941 | matchAndRewrite(fir::LoadOp load, OpAdaptor adaptor, | ||||||
| 2942 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2943 | if (auto boxTy = load.getType().dyn_cast<fir::BaseBoxType>()) { | ||||||
| 2944 | // fir.box is a special case because it is considered as an ssa values in | ||||||
| 2945 | // fir, but it is lowered as a pointer to a descriptor. So | ||||||
| 2946 | // fir.ref<fir.box> and fir.box end up being the same llvm types and | ||||||
| 2947 | // loading a fir.ref<fir.box> is implemented as taking a snapshot of the | ||||||
| 2948 | // descriptor value into a new descriptor temp. | ||||||
| 2949 | auto inputBoxStorage = adaptor.getOperands()[0]; | ||||||
| 2950 | mlir::Location loc = load.getLoc(); | ||||||
| 2951 | fir::SequenceType seqTy = fir::unwrapUntilSeqType(boxTy); | ||||||
| 2952 | // fir.box of assumed rank do not have a storage | ||||||
| 2953 | // size that is know at compile time. The copy needs to be runtime driven | ||||||
| 2954 | // depending on the actual dynamic rank or type. | ||||||
| 2955 | if (seqTy && seqTy.hasUnknownShape()) | ||||||
| 2956 | TODO(loc, "loading or assumed rank fir.box")do { fir::emitFatalError(loc, llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "2956" ": not yet implemented: ") + llvm::Twine("loading or assumed rank fir.box" ), false); } while (false); | ||||||
| 2957 | mlir::Type boxPtrTy = inputBoxStorage.getType(); | ||||||
| 2958 | auto boxValue = rewriter.create<mlir::LLVM::LoadOp>( | ||||||
| 2959 | loc, boxPtrTy.cast<mlir::LLVM::LLVMPointerType>().getElementType(), | ||||||
| 2960 | inputBoxStorage); | ||||||
| 2961 | attachTBAATag(boxValue, boxTy, boxTy, nullptr); | ||||||
| 2962 | auto newBoxStorage = | ||||||
| 2963 | genAllocaWithType(loc, boxPtrTy, defaultAlign, rewriter); | ||||||
| 2964 | auto storeOp = | ||||||
| 2965 | rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, newBoxStorage); | ||||||
| 2966 | attachTBAATag(storeOp, boxTy, boxTy, nullptr); | ||||||
| 2967 | rewriter.replaceOp(load, newBoxStorage.getResult()); | ||||||
| 2968 | } else { | ||||||
| 2969 | mlir::Type loadTy = convertType(load.getType()); | ||||||
| 2970 | auto loadOp = rewriter.create<mlir::LLVM::LoadOp>( | ||||||
| 2971 | load.getLoc(), loadTy, adaptor.getOperands(), load->getAttrs()); | ||||||
| 2972 | attachTBAATag(loadOp, load.getType(), load.getType(), nullptr); | ||||||
| 2973 | rewriter.replaceOp(load, loadOp.getResult()); | ||||||
| 2974 | } | ||||||
| 2975 | return mlir::success(); | ||||||
| 2976 | } | ||||||
| 2977 | }; | ||||||
| 2978 | |||||||
| 2979 | /// Lower `fir.no_reassoc` to LLVM IR dialect. | ||||||
| 2980 | /// TODO: how do we want to enforce this in LLVM-IR? Can we manipulate the fast | ||||||
| 2981 | /// math flags? | ||||||
| 2982 | struct NoReassocOpConversion : public FIROpConversion<fir::NoReassocOp> { | ||||||
| 2983 | using FIROpConversion::FIROpConversion; | ||||||
| 2984 | |||||||
| 2985 | mlir::LogicalResult | ||||||
| 2986 | matchAndRewrite(fir::NoReassocOp noreassoc, OpAdaptor adaptor, | ||||||
| 2987 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 2988 | rewriter.replaceOp(noreassoc, adaptor.getOperands()[0]); | ||||||
| 2989 | return mlir::success(); | ||||||
| 2990 | } | ||||||
| 2991 | }; | ||||||
| 2992 | |||||||
| 2993 | static void genCondBrOp(mlir::Location loc, mlir::Value cmp, mlir::Block *dest, | ||||||
| 2994 | std::optional<mlir::ValueRange> destOps, | ||||||
| 2995 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 2996 | mlir::Block *newBlock) { | ||||||
| 2997 | if (destOps) | ||||||
| 2998 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, dest, *destOps, newBlock, | ||||||
| 2999 | mlir::ValueRange()); | ||||||
| 3000 | else | ||||||
| 3001 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, dest, newBlock); | ||||||
| 3002 | } | ||||||
| 3003 | |||||||
| 3004 | template <typename A, typename B> | ||||||
| 3005 | static void genBrOp(A caseOp, mlir::Block *dest, std::optional<B> destOps, | ||||||
| 3006 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 3007 | if (destOps) | ||||||
| 3008 | rewriter.replaceOpWithNewOp<mlir::LLVM::BrOp>(caseOp, *destOps, dest); | ||||||
| 3009 | else | ||||||
| 3010 | rewriter.replaceOpWithNewOp<mlir::LLVM::BrOp>(caseOp, std::nullopt, dest); | ||||||
| 3011 | } | ||||||
| 3012 | |||||||
| 3013 | static void genCaseLadderStep(mlir::Location loc, mlir::Value cmp, | ||||||
| 3014 | mlir::Block *dest, | ||||||
| 3015 | std::optional<mlir::ValueRange> destOps, | ||||||
| 3016 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 3017 | auto *thisBlock = rewriter.getInsertionBlock(); | ||||||
| 3018 | auto *newBlock = createBlock(rewriter, dest); | ||||||
| 3019 | rewriter.setInsertionPointToEnd(thisBlock); | ||||||
| 3020 | genCondBrOp(loc, cmp, dest, destOps, rewriter, newBlock); | ||||||
| 3021 | rewriter.setInsertionPointToEnd(newBlock); | ||||||
| 3022 | } | ||||||
| 3023 | |||||||
| 3024 | /// Conversion of `fir.select_case` | ||||||
| 3025 | /// | ||||||
| 3026 | /// The `fir.select_case` operation is converted to a if-then-else ladder. | ||||||
| 3027 | /// Depending on the case condition type, one or several comparison and | ||||||
| 3028 | /// conditional branching can be generated. | ||||||
| 3029 | /// | ||||||
| 3030 | /// A a point value case such as `case(4)`, a lower bound case such as | ||||||
| 3031 | /// `case(5:)` or an upper bound case such as `case(:3)` are converted to a | ||||||
| 3032 | /// simple comparison between the selector value and the constant value in the | ||||||
| 3033 | /// case. The block associated with the case condition is then executed if | ||||||
| 3034 | /// the comparison succeed otherwise it branch to the next block with the | ||||||
| 3035 | /// comparison for the the next case conditon. | ||||||
| 3036 | /// | ||||||
| 3037 | /// A closed interval case condition such as `case(7:10)` is converted with a | ||||||
| 3038 | /// first comparison and conditional branching for the lower bound. If | ||||||
| 3039 | /// successful, it branch to a second block with the comparison for the | ||||||
| 3040 | /// upper bound in the same case condition. | ||||||
| 3041 | /// | ||||||
| 3042 | /// TODO: lowering of CHARACTER type cases is not handled yet. | ||||||
| 3043 | struct SelectCaseOpConversion : public FIROpConversion<fir::SelectCaseOp> { | ||||||
| 3044 | using FIROpConversion::FIROpConversion; | ||||||
| 3045 | |||||||
| 3046 | mlir::LogicalResult | ||||||
| 3047 | matchAndRewrite(fir::SelectCaseOp caseOp, OpAdaptor adaptor, | ||||||
| 3048 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3049 | unsigned conds = caseOp.getNumConditions(); | ||||||
| 3050 | llvm::ArrayRef<mlir::Attribute> cases = caseOp.getCases().getValue(); | ||||||
| 3051 | // Type can be CHARACTER, INTEGER, or LOGICAL (C1145) | ||||||
| 3052 | auto ty = caseOp.getSelector().getType(); | ||||||
| 3053 | if (ty.isa<fir::CharacterType>()) { | ||||||
| 3054 | TODO(caseOp.getLoc(), "fir.select_case codegen with character type")do { fir::emitFatalError(caseOp.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "3054" ": not yet implemented: ") + llvm::Twine("fir.select_case codegen with character type" ), false); } while (false); | ||||||
| 3055 | return mlir::failure(); | ||||||
| 3056 | } | ||||||
| 3057 | mlir::Value selector = caseOp.getSelector(adaptor.getOperands()); | ||||||
| 3058 | auto loc = caseOp.getLoc(); | ||||||
| 3059 | for (unsigned t = 0; t != conds; ++t) { | ||||||
| 3060 | mlir::Block *dest = caseOp.getSuccessor(t); | ||||||
| 3061 | std::optional<mlir::ValueRange> destOps = | ||||||
| 3062 | caseOp.getSuccessorOperands(adaptor.getOperands(), t); | ||||||
| 3063 | std::optional<mlir::ValueRange> cmpOps = | ||||||
| 3064 | *caseOp.getCompareOperands(adaptor.getOperands(), t); | ||||||
| 3065 | mlir::Value caseArg = *(cmpOps.value().begin()); | ||||||
| 3066 | mlir::Attribute attr = cases[t]; | ||||||
| 3067 | if (attr.isa<fir::PointIntervalAttr>()) { | ||||||
| 3068 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 3069 | loc, mlir::LLVM::ICmpPredicate::eq, selector, caseArg); | ||||||
| 3070 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); | ||||||
| 3071 | continue; | ||||||
| 3072 | } | ||||||
| 3073 | if (attr.isa<fir::LowerBoundAttr>()) { | ||||||
| 3074 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 3075 | loc, mlir::LLVM::ICmpPredicate::sle, caseArg, selector); | ||||||
| 3076 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); | ||||||
| 3077 | continue; | ||||||
| 3078 | } | ||||||
| 3079 | if (attr.isa<fir::UpperBoundAttr>()) { | ||||||
| 3080 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 3081 | loc, mlir::LLVM::ICmpPredicate::sle, selector, caseArg); | ||||||
| 3082 | genCaseLadderStep(loc, cmp, dest, destOps, rewriter); | ||||||
| 3083 | continue; | ||||||
| 3084 | } | ||||||
| 3085 | if (attr.isa<fir::ClosedIntervalAttr>()) { | ||||||
| 3086 | auto cmp = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 3087 | loc, mlir::LLVM::ICmpPredicate::sle, caseArg, selector); | ||||||
| 3088 | auto *thisBlock = rewriter.getInsertionBlock(); | ||||||
| 3089 | auto *newBlock1 = createBlock(rewriter, dest); | ||||||
| 3090 | auto *newBlock2 = createBlock(rewriter, dest); | ||||||
| 3091 | rewriter.setInsertionPointToEnd(thisBlock); | ||||||
| 3092 | rewriter.create<mlir::LLVM::CondBrOp>(loc, cmp, newBlock1, newBlock2); | ||||||
| 3093 | rewriter.setInsertionPointToEnd(newBlock1); | ||||||
| 3094 | mlir::Value caseArg0 = *(cmpOps.value().begin() + 1); | ||||||
| 3095 | auto cmp0 = rewriter.create<mlir::LLVM::ICmpOp>( | ||||||
| 3096 | loc, mlir::LLVM::ICmpPredicate::sle, selector, caseArg0); | ||||||
| 3097 | genCondBrOp(loc, cmp0, dest, destOps, rewriter, newBlock2); | ||||||
| 3098 | rewriter.setInsertionPointToEnd(newBlock2); | ||||||
| 3099 | continue; | ||||||
| 3100 | } | ||||||
| 3101 | assert(attr.isa<mlir::UnitAttr>())(static_cast <bool> (attr.isa<mlir::UnitAttr>()) ? void (0) : __assert_fail ("attr.isa<mlir::UnitAttr>()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3101, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3102 | assert((t + 1 == conds) && "unit must be last")(static_cast <bool> ((t + 1 == conds) && "unit must be last" ) ? void (0) : __assert_fail ("(t + 1 == conds) && \"unit must be last\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3102, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3103 | genBrOp(caseOp, dest, destOps, rewriter); | ||||||
| 3104 | } | ||||||
| 3105 | return mlir::success(); | ||||||
| 3106 | } | ||||||
| 3107 | }; | ||||||
| 3108 | |||||||
| 3109 | template <typename OP> | ||||||
| 3110 | static void selectMatchAndRewrite(fir::LLVMTypeConverter &lowering, OP select, | ||||||
| 3111 | typename OP::Adaptor adaptor, | ||||||
| 3112 | mlir::ConversionPatternRewriter &rewriter) { | ||||||
| 3113 | unsigned conds = select.getNumConditions(); | ||||||
| 3114 | auto cases = select.getCases().getValue(); | ||||||
| 3115 | mlir::Value selector = adaptor.getSelector(); | ||||||
| 3116 | auto loc = select.getLoc(); | ||||||
| 3117 | assert(conds > 0 && "select must have cases")(static_cast <bool> (conds > 0 && "select must have cases" ) ? void (0) : __assert_fail ("conds > 0 && \"select must have cases\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3117, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3118 | |||||||
| 3119 | llvm::SmallVector<mlir::Block *> destinations; | ||||||
| 3120 | llvm::SmallVector<mlir::ValueRange> destinationsOperands; | ||||||
| 3121 | mlir::Block *defaultDestination; | ||||||
| 3122 | mlir::ValueRange defaultOperands; | ||||||
| 3123 | llvm::SmallVector<int32_t> caseValues; | ||||||
| 3124 | |||||||
| 3125 | for (unsigned t = 0; t
| ||||||
| 3126 | mlir::Block *dest = select.getSuccessor(t); | ||||||
| 3127 | auto destOps = select.getSuccessorOperands(adaptor.getOperands(), t); | ||||||
| 3128 | const mlir::Attribute &attr = cases[t]; | ||||||
| 3129 | if (auto intAttr = attr.template dyn_cast<mlir::IntegerAttr>()) { | ||||||
| 3130 | destinations.push_back(dest); | ||||||
| 3131 | destinationsOperands.push_back(destOps ? *destOps : mlir::ValueRange{}); | ||||||
| 3132 | caseValues.push_back(intAttr.getInt()); | ||||||
| 3133 | continue; | ||||||
| 3134 | } | ||||||
| 3135 | assert(attr.template dyn_cast_or_null<mlir::UnitAttr>())(static_cast <bool> (attr.template dyn_cast_or_null< mlir::UnitAttr>()) ? void (0) : __assert_fail ("attr.template dyn_cast_or_null<mlir::UnitAttr>()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3135, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3136 | assert((t + 1 == conds) && "unit must be last")(static_cast <bool> ((t + 1 == conds) && "unit must be last" ) ? void (0) : __assert_fail ("(t + 1 == conds) && \"unit must be last\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3136, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3137 | defaultDestination = dest; | ||||||
| 3138 | defaultOperands = destOps ? *destOps : mlir::ValueRange{}; | ||||||
| 3139 | } | ||||||
| 3140 | |||||||
| 3141 | // LLVM::SwitchOp takes a i32 type for the selector. | ||||||
| 3142 | if (select.getSelector().getType() != rewriter.getI32Type()) | ||||||
| 3143 | selector = rewriter.create<mlir::LLVM::TruncOp>(loc, rewriter.getI32Type(), | ||||||
| 3144 | selector); | ||||||
| 3145 | |||||||
| 3146 | rewriter.replaceOpWithNewOp<mlir::LLVM::SwitchOp>( | ||||||
| 3147 | select, selector, | ||||||
| 3148 | /*defaultDestination=*/defaultDestination, | ||||||
| 3149 | /*defaultOperands=*/defaultOperands, | ||||||
| 3150 | /*caseValues=*/caseValues, | ||||||
| 3151 | /*caseDestinations=*/destinations, | ||||||
| 3152 | /*caseOperands=*/destinationsOperands, | ||||||
| 3153 | /*branchWeights=*/llvm::ArrayRef<std::int32_t>()); | ||||||
| 3154 | } | ||||||
| 3155 | |||||||
| 3156 | /// conversion of fir::SelectOp to an if-then-else ladder | ||||||
| 3157 | struct SelectOpConversion : public FIROpConversion<fir::SelectOp> { | ||||||
| 3158 | using FIROpConversion::FIROpConversion; | ||||||
| 3159 | |||||||
| 3160 | mlir::LogicalResult | ||||||
| 3161 | matchAndRewrite(fir::SelectOp op, OpAdaptor adaptor, | ||||||
| 3162 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3163 | selectMatchAndRewrite<fir::SelectOp>(lowerTy(), op, adaptor, rewriter); | ||||||
| 3164 | return mlir::success(); | ||||||
| 3165 | } | ||||||
| 3166 | }; | ||||||
| 3167 | |||||||
| 3168 | /// conversion of fir::SelectRankOp to an if-then-else ladder | ||||||
| 3169 | struct SelectRankOpConversion : public FIROpConversion<fir::SelectRankOp> { | ||||||
| 3170 | using FIROpConversion::FIROpConversion; | ||||||
| 3171 | |||||||
| 3172 | mlir::LogicalResult | ||||||
| 3173 | matchAndRewrite(fir::SelectRankOp op, OpAdaptor adaptor, | ||||||
| 3174 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3175 | selectMatchAndRewrite<fir::SelectRankOp>(lowerTy(), op, adaptor, rewriter); | ||||||
| |||||||
| 3176 | return mlir::success(); | ||||||
| 3177 | } | ||||||
| 3178 | }; | ||||||
| 3179 | |||||||
| 3180 | /// Lower `fir.select_type` to LLVM IR dialect. | ||||||
| 3181 | struct SelectTypeOpConversion : public FIROpConversion<fir::SelectTypeOp> { | ||||||
| 3182 | using FIROpConversion::FIROpConversion; | ||||||
| 3183 | |||||||
| 3184 | mlir::LogicalResult | ||||||
| 3185 | matchAndRewrite(fir::SelectTypeOp select, OpAdaptor adaptor, | ||||||
| 3186 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3187 | mlir::emitError(select.getLoc(), | ||||||
| 3188 | "fir.select_type should have already been converted"); | ||||||
| 3189 | return mlir::failure(); | ||||||
| 3190 | } | ||||||
| 3191 | }; | ||||||
| 3192 | |||||||
| 3193 | /// `fir.store` --> `llvm.store` | ||||||
| 3194 | struct StoreOpConversion : public FIROpConversion<fir::StoreOp> { | ||||||
| 3195 | using FIROpConversion::FIROpConversion; | ||||||
| 3196 | |||||||
| 3197 | mlir::LogicalResult | ||||||
| 3198 | matchAndRewrite(fir::StoreOp store, OpAdaptor adaptor, | ||||||
| 3199 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3200 | mlir::Location loc = store.getLoc(); | ||||||
| 3201 | mlir::Type storeTy = store.getValue().getType(); | ||||||
| 3202 | mlir::LLVM::StoreOp newStoreOp; | ||||||
| 3203 | if (auto boxTy = storeTy.dyn_cast<fir::BaseBoxType>()) { | ||||||
| 3204 | // fir.box value is actually in memory, load it first before storing it. | ||||||
| 3205 | mlir::Type boxPtrTy = adaptor.getOperands()[0].getType(); | ||||||
| 3206 | auto val = rewriter.create<mlir::LLVM::LoadOp>( | ||||||
| 3207 | loc, boxPtrTy.cast<mlir::LLVM::LLVMPointerType>().getElementType(), | ||||||
| 3208 | adaptor.getOperands()[0]); | ||||||
| 3209 | attachTBAATag(val, boxTy, boxTy, nullptr); | ||||||
| 3210 | newStoreOp = rewriter.create<mlir::LLVM::StoreOp>( | ||||||
| 3211 | loc, val, adaptor.getOperands()[1]); | ||||||
| 3212 | } else { | ||||||
| 3213 | newStoreOp = rewriter.create<mlir::LLVM::StoreOp>( | ||||||
| 3214 | loc, adaptor.getOperands()[0], adaptor.getOperands()[1]); | ||||||
| 3215 | } | ||||||
| 3216 | attachTBAATag(newStoreOp, storeTy, storeTy, nullptr); | ||||||
| 3217 | rewriter.eraseOp(store); | ||||||
| 3218 | return mlir::success(); | ||||||
| 3219 | } | ||||||
| 3220 | }; | ||||||
| 3221 | |||||||
| 3222 | namespace { | ||||||
| 3223 | |||||||
| 3224 | /// Convert `fir.unboxchar` into two `llvm.extractvalue` instructions. One for | ||||||
| 3225 | /// the character buffer and one for the buffer length. | ||||||
| 3226 | struct UnboxCharOpConversion : public FIROpConversion<fir::UnboxCharOp> { | ||||||
| 3227 | using FIROpConversion::FIROpConversion; | ||||||
| 3228 | |||||||
| 3229 | mlir::LogicalResult | ||||||
| 3230 | matchAndRewrite(fir::UnboxCharOp unboxchar, OpAdaptor adaptor, | ||||||
| 3231 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3232 | mlir::Type lenTy = convertType(unboxchar.getType(1)); | ||||||
| 3233 | mlir::Value tuple = adaptor.getOperands()[0]; | ||||||
| 3234 | |||||||
| 3235 | mlir::Location loc = unboxchar.getLoc(); | ||||||
| 3236 | mlir::Value ptrToBuffer = | ||||||
| 3237 | rewriter.create<mlir::LLVM::ExtractValueOp>(loc, tuple, 0); | ||||||
| 3238 | |||||||
| 3239 | auto len = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, tuple, 1); | ||||||
| 3240 | mlir::Value lenAfterCast = integerCast(loc, rewriter, lenTy, len); | ||||||
| 3241 | |||||||
| 3242 | rewriter.replaceOp(unboxchar, | ||||||
| 3243 | llvm::ArrayRef<mlir::Value>{ptrToBuffer, lenAfterCast}); | ||||||
| 3244 | return mlir::success(); | ||||||
| 3245 | } | ||||||
| 3246 | }; | ||||||
| 3247 | |||||||
| 3248 | /// Lower `fir.unboxproc` operation. Unbox a procedure box value, yielding its | ||||||
| 3249 | /// components. | ||||||
| 3250 | /// TODO: Part of supporting Fortran 2003 procedure pointers. | ||||||
| 3251 | struct UnboxProcOpConversion : public FIROpConversion<fir::UnboxProcOp> { | ||||||
| 3252 | using FIROpConversion::FIROpConversion; | ||||||
| 3253 | |||||||
| 3254 | mlir::LogicalResult | ||||||
| 3255 | matchAndRewrite(fir::UnboxProcOp unboxproc, OpAdaptor adaptor, | ||||||
| 3256 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3257 | TODO(unboxproc.getLoc(), "fir.unboxproc codegen")do { fir::emitFatalError(unboxproc.getLoc(), llvm::Twine("flang/lib/Optimizer/CodeGen/CodeGen.cpp" ":" "3257" ": not yet implemented: ") + llvm::Twine("fir.unboxproc codegen" ), false); } while (false); | ||||||
| 3258 | return mlir::failure(); | ||||||
| 3259 | } | ||||||
| 3260 | }; | ||||||
| 3261 | |||||||
| 3262 | /// convert to LLVM IR dialect `undef` | ||||||
| 3263 | struct UndefOpConversion : public FIROpConversion<fir::UndefOp> { | ||||||
| 3264 | using FIROpConversion::FIROpConversion; | ||||||
| 3265 | |||||||
| 3266 | mlir::LogicalResult | ||||||
| 3267 | matchAndRewrite(fir::UndefOp undef, OpAdaptor, | ||||||
| 3268 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3269 | rewriter.replaceOpWithNewOp<mlir::LLVM::UndefOp>( | ||||||
| 3270 | undef, convertType(undef.getType())); | ||||||
| 3271 | return mlir::success(); | ||||||
| 3272 | } | ||||||
| 3273 | }; | ||||||
| 3274 | |||||||
| 3275 | struct ZeroOpConversion : public FIROpConversion<fir::ZeroOp> { | ||||||
| 3276 | using FIROpConversion::FIROpConversion; | ||||||
| 3277 | |||||||
| 3278 | mlir::LogicalResult | ||||||
| 3279 | matchAndRewrite(fir::ZeroOp zero, OpAdaptor, | ||||||
| 3280 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3281 | mlir::Type ty = convertType(zero.getType()); | ||||||
| 3282 | if (ty.isa<mlir::LLVM::LLVMPointerType>()) { | ||||||
| 3283 | rewriter.replaceOpWithNewOp<mlir::LLVM::NullOp>(zero, ty); | ||||||
| 3284 | } else if (ty.isa<mlir::IntegerType>()) { | ||||||
| 3285 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( | ||||||
| 3286 | zero, ty, mlir::IntegerAttr::get(ty, 0)); | ||||||
| 3287 | } else if (mlir::LLVM::isCompatibleFloatingPointType(ty)) { | ||||||
| 3288 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( | ||||||
| 3289 | zero, ty, mlir::FloatAttr::get(zero.getType(), 0.0)); | ||||||
| 3290 | } else { | ||||||
| 3291 | // TODO: create ConstantAggregateZero for FIR aggregate/array types. | ||||||
| 3292 | return rewriter.notifyMatchFailure( | ||||||
| 3293 | zero, | ||||||
| 3294 | "conversion of fir.zero with aggregate type not implemented yet"); | ||||||
| 3295 | } | ||||||
| 3296 | return mlir::success(); | ||||||
| 3297 | } | ||||||
| 3298 | }; | ||||||
| 3299 | |||||||
| 3300 | /// `fir.unreachable` --> `llvm.unreachable` | ||||||
| 3301 | struct UnreachableOpConversion : public FIROpConversion<fir::UnreachableOp> { | ||||||
| 3302 | using FIROpConversion::FIROpConversion; | ||||||
| 3303 | |||||||
| 3304 | mlir::LogicalResult | ||||||
| 3305 | matchAndRewrite(fir::UnreachableOp unreach, OpAdaptor adaptor, | ||||||
| 3306 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3307 | rewriter.replaceOpWithNewOp<mlir::LLVM::UnreachableOp>(unreach); | ||||||
| 3308 | return mlir::success(); | ||||||
| 3309 | } | ||||||
| 3310 | }; | ||||||
| 3311 | |||||||
| 3312 | /// `fir.is_present` --> | ||||||
| 3313 | /// ``` | ||||||
| 3314 | /// %0 = llvm.mlir.constant(0 : i64) | ||||||
| 3315 | /// %1 = llvm.ptrtoint %0 | ||||||
| 3316 | /// %2 = llvm.icmp "ne" %1, %0 : i64 | ||||||
| 3317 | /// ``` | ||||||
| 3318 | struct IsPresentOpConversion : public FIROpConversion<fir::IsPresentOp> { | ||||||
| 3319 | using FIROpConversion::FIROpConversion; | ||||||
| 3320 | |||||||
| 3321 | mlir::LogicalResult | ||||||
| 3322 | matchAndRewrite(fir::IsPresentOp isPresent, OpAdaptor adaptor, | ||||||
| 3323 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3324 | mlir::Type idxTy = lowerTy().indexType(); | ||||||
| 3325 | mlir::Location loc = isPresent.getLoc(); | ||||||
| 3326 | auto ptr = adaptor.getOperands()[0]; | ||||||
| 3327 | |||||||
| 3328 | if (isPresent.getVal().getType().isa<fir::BoxCharType>()) { | ||||||
| 3329 | [[maybe_unused]] auto structTy = | ||||||
| 3330 | ptr.getType().cast<mlir::LLVM::LLVMStructType>(); | ||||||
| 3331 | assert(!structTy.isOpaque() && !structTy.getBody().empty())(static_cast <bool> (!structTy.isOpaque() && !structTy .getBody().empty()) ? void (0) : __assert_fail ("!structTy.isOpaque() && !structTy.getBody().empty()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3331, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3332 | |||||||
| 3333 | ptr = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, ptr, 0); | ||||||
| 3334 | } | ||||||
| 3335 | mlir::LLVM::ConstantOp c0 = | ||||||
| 3336 | genConstantIndex(isPresent.getLoc(), idxTy, rewriter, 0); | ||||||
| 3337 | auto addr = rewriter.create<mlir::LLVM::PtrToIntOp>(loc, idxTy, ptr); | ||||||
| 3338 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( | ||||||
| 3339 | isPresent, mlir::LLVM::ICmpPredicate::ne, addr, c0); | ||||||
| 3340 | |||||||
| 3341 | return mlir::success(); | ||||||
| 3342 | } | ||||||
| 3343 | }; | ||||||
| 3344 | |||||||
| 3345 | /// Create value signaling an absent optional argument in a call, e.g. | ||||||
| 3346 | /// `fir.absent !fir.ref<i64>` --> `llvm.mlir.null : !llvm.ptr<i64>` | ||||||
| 3347 | struct AbsentOpConversion : public FIROpConversion<fir::AbsentOp> { | ||||||
| 3348 | using FIROpConversion::FIROpConversion; | ||||||
| 3349 | |||||||
| 3350 | mlir::LogicalResult | ||||||
| 3351 | matchAndRewrite(fir::AbsentOp absent, OpAdaptor, | ||||||
| 3352 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3353 | mlir::Type ty = convertType(absent.getType()); | ||||||
| 3354 | mlir::Location loc = absent.getLoc(); | ||||||
| 3355 | |||||||
| 3356 | if (absent.getType().isa<fir::BoxCharType>()) { | ||||||
| 3357 | auto structTy = ty.cast<mlir::LLVM::LLVMStructType>(); | ||||||
| 3358 | assert(!structTy.isOpaque() && !structTy.getBody().empty())(static_cast <bool> (!structTy.isOpaque() && !structTy .getBody().empty()) ? void (0) : __assert_fail ("!structTy.isOpaque() && !structTy.getBody().empty()" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3358, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3359 | auto undefStruct = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 3360 | auto nullField = | ||||||
| 3361 | rewriter.create<mlir::LLVM::NullOp>(loc, structTy.getBody()[0]); | ||||||
| 3362 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( | ||||||
| 3363 | absent, undefStruct, nullField, 0); | ||||||
| 3364 | } else { | ||||||
| 3365 | rewriter.replaceOpWithNewOp<mlir::LLVM::NullOp>(absent, ty); | ||||||
| 3366 | } | ||||||
| 3367 | return mlir::success(); | ||||||
| 3368 | } | ||||||
| 3369 | }; | ||||||
| 3370 | |||||||
| 3371 | // | ||||||
| 3372 | // Primitive operations on Complex types | ||||||
| 3373 | // | ||||||
| 3374 | |||||||
| 3375 | /// Generate inline code for complex addition/subtraction | ||||||
| 3376 | template <typename LLVMOP, typename OPTY> | ||||||
| 3377 | static mlir::LLVM::InsertValueOp | ||||||
| 3378 | complexSum(OPTY sumop, mlir::ValueRange opnds, | ||||||
| 3379 | mlir::ConversionPatternRewriter &rewriter, | ||||||
| 3380 | fir::LLVMTypeConverter &lowering) { | ||||||
| 3381 | mlir::Value a = opnds[0]; | ||||||
| 3382 | mlir::Value b = opnds[1]; | ||||||
| 3383 | auto loc = sumop.getLoc(); | ||||||
| 3384 | mlir::Type eleTy = lowering.convertType(getComplexEleTy(sumop.getType())); | ||||||
| 3385 | mlir::Type ty = lowering.convertType(sumop.getType()); | ||||||
| 3386 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); | ||||||
| 3387 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); | ||||||
| 3388 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); | ||||||
| 3389 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); | ||||||
| 3390 | auto rx = rewriter.create<LLVMOP>(loc, eleTy, x0, x1); | ||||||
| 3391 | auto ry = rewriter.create<LLVMOP>(loc, eleTy, y0, y1); | ||||||
| 3392 | auto r0 = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 3393 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r0, rx, 0); | ||||||
| 3394 | return rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ry, 1); | ||||||
| 3395 | } | ||||||
| 3396 | } // namespace | ||||||
| 3397 | |||||||
| 3398 | namespace { | ||||||
| 3399 | struct AddcOpConversion : public FIROpConversion<fir::AddcOp> { | ||||||
| 3400 | using FIROpConversion::FIROpConversion; | ||||||
| 3401 | |||||||
| 3402 | mlir::LogicalResult | ||||||
| 3403 | matchAndRewrite(fir::AddcOp addc, OpAdaptor adaptor, | ||||||
| 3404 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3405 | // given: (x + iy) + (x' + iy') | ||||||
| 3406 | // result: (x + x') + i(y + y') | ||||||
| 3407 | auto r = complexSum<mlir::LLVM::FAddOp>(addc, adaptor.getOperands(), | ||||||
| 3408 | rewriter, lowerTy()); | ||||||
| 3409 | rewriter.replaceOp(addc, r.getResult()); | ||||||
| 3410 | return mlir::success(); | ||||||
| 3411 | } | ||||||
| 3412 | }; | ||||||
| 3413 | |||||||
| 3414 | struct SubcOpConversion : public FIROpConversion<fir::SubcOp> { | ||||||
| 3415 | using FIROpConversion::FIROpConversion; | ||||||
| 3416 | |||||||
| 3417 | mlir::LogicalResult | ||||||
| 3418 | matchAndRewrite(fir::SubcOp subc, OpAdaptor adaptor, | ||||||
| 3419 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3420 | // given: (x + iy) - (x' + iy') | ||||||
| 3421 | // result: (x - x') + i(y - y') | ||||||
| 3422 | auto r = complexSum<mlir::LLVM::FSubOp>(subc, adaptor.getOperands(), | ||||||
| 3423 | rewriter, lowerTy()); | ||||||
| 3424 | rewriter.replaceOp(subc, r.getResult()); | ||||||
| 3425 | return mlir::success(); | ||||||
| 3426 | } | ||||||
| 3427 | }; | ||||||
| 3428 | |||||||
| 3429 | /// Inlined complex multiply | ||||||
| 3430 | struct MulcOpConversion : public FIROpConversion<fir::MulcOp> { | ||||||
| 3431 | using FIROpConversion::FIROpConversion; | ||||||
| 3432 | |||||||
| 3433 | mlir::LogicalResult | ||||||
| 3434 | matchAndRewrite(fir::MulcOp mulc, OpAdaptor adaptor, | ||||||
| 3435 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3436 | // TODO: Can we use a call to __muldc3 ? | ||||||
| 3437 | // given: (x + iy) * (x' + iy') | ||||||
| 3438 | // result: (xx'-yy')+i(xy'+yx') | ||||||
| 3439 | mlir::Value a = adaptor.getOperands()[0]; | ||||||
| 3440 | mlir::Value b = adaptor.getOperands()[1]; | ||||||
| 3441 | auto loc = mulc.getLoc(); | ||||||
| 3442 | mlir::Type eleTy = convertType(getComplexEleTy(mulc.getType())); | ||||||
| 3443 | mlir::Type ty = convertType(mulc.getType()); | ||||||
| 3444 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); | ||||||
| 3445 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); | ||||||
| 3446 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); | ||||||
| 3447 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); | ||||||
| 3448 | auto xx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, x1); | ||||||
| 3449 | auto yx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, x1); | ||||||
| 3450 | auto xy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, y1); | ||||||
| 3451 | auto ri = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, xy, yx); | ||||||
| 3452 | auto yy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, y1); | ||||||
| 3453 | auto rr = rewriter.create<mlir::LLVM::FSubOp>(loc, eleTy, xx, yy); | ||||||
| 3454 | auto ra = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 3455 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, ra, rr, 0); | ||||||
| 3456 | auto r0 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ri, 1); | ||||||
| 3457 | rewriter.replaceOp(mulc, r0.getResult()); | ||||||
| 3458 | return mlir::success(); | ||||||
| 3459 | } | ||||||
| 3460 | }; | ||||||
| 3461 | |||||||
| 3462 | /// Inlined complex division | ||||||
| 3463 | struct DivcOpConversion : public FIROpConversion<fir::DivcOp> { | ||||||
| 3464 | using FIROpConversion::FIROpConversion; | ||||||
| 3465 | |||||||
| 3466 | mlir::LogicalResult | ||||||
| 3467 | matchAndRewrite(fir::DivcOp divc, OpAdaptor adaptor, | ||||||
| 3468 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3469 | // TODO: Can we use a call to __divdc3 instead? | ||||||
| 3470 | // Just generate inline code for now. | ||||||
| 3471 | // given: (x + iy) / (x' + iy') | ||||||
| 3472 | // result: ((xx'+yy')/d) + i((yx'-xy')/d) where d = x'x' + y'y' | ||||||
| 3473 | mlir::Value a = adaptor.getOperands()[0]; | ||||||
| 3474 | mlir::Value b = adaptor.getOperands()[1]; | ||||||
| 3475 | auto loc = divc.getLoc(); | ||||||
| 3476 | mlir::Type eleTy = convertType(getComplexEleTy(divc.getType())); | ||||||
| 3477 | mlir::Type ty = convertType(divc.getType()); | ||||||
| 3478 | auto x0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 0); | ||||||
| 3479 | auto y0 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, a, 1); | ||||||
| 3480 | auto x1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 0); | ||||||
| 3481 | auto y1 = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, b, 1); | ||||||
| 3482 | auto xx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, x1); | ||||||
| 3483 | auto x1x1 = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x1, x1); | ||||||
| 3484 | auto yx = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, x1); | ||||||
| 3485 | auto xy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, x0, y1); | ||||||
| 3486 | auto yy = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y0, y1); | ||||||
| 3487 | auto y1y1 = rewriter.create<mlir::LLVM::FMulOp>(loc, eleTy, y1, y1); | ||||||
| 3488 | auto d = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, x1x1, y1y1); | ||||||
| 3489 | auto rrn = rewriter.create<mlir::LLVM::FAddOp>(loc, eleTy, xx, yy); | ||||||
| 3490 | auto rin = rewriter.create<mlir::LLVM::FSubOp>(loc, eleTy, yx, xy); | ||||||
| 3491 | auto rr = rewriter.create<mlir::LLVM::FDivOp>(loc, eleTy, rrn, d); | ||||||
| 3492 | auto ri = rewriter.create<mlir::LLVM::FDivOp>(loc, eleTy, rin, d); | ||||||
| 3493 | auto ra = rewriter.create<mlir::LLVM::UndefOp>(loc, ty); | ||||||
| 3494 | auto r1 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, ra, rr, 0); | ||||||
| 3495 | auto r0 = rewriter.create<mlir::LLVM::InsertValueOp>(loc, r1, ri, 1); | ||||||
| 3496 | rewriter.replaceOp(divc, r0.getResult()); | ||||||
| 3497 | return mlir::success(); | ||||||
| 3498 | } | ||||||
| 3499 | }; | ||||||
| 3500 | |||||||
| 3501 | /// Inlined complex negation | ||||||
| 3502 | struct NegcOpConversion : public FIROpConversion<fir::NegcOp> { | ||||||
| 3503 | using FIROpConversion::FIROpConversion; | ||||||
| 3504 | |||||||
| 3505 | mlir::LogicalResult | ||||||
| 3506 | matchAndRewrite(fir::NegcOp neg, OpAdaptor adaptor, | ||||||
| 3507 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3508 | // given: -(x + iy) | ||||||
| 3509 | // result: -x - iy | ||||||
| 3510 | auto eleTy = convertType(getComplexEleTy(neg.getType())); | ||||||
| 3511 | auto loc = neg.getLoc(); | ||||||
| 3512 | mlir::Value o0 = adaptor.getOperands()[0]; | ||||||
| 3513 | auto rp = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, o0, 0); | ||||||
| 3514 | auto ip = rewriter.create<mlir::LLVM::ExtractValueOp>(loc, o0, 1); | ||||||
| 3515 | auto nrp = rewriter.create<mlir::LLVM::FNegOp>(loc, eleTy, rp); | ||||||
| 3516 | auto nip = rewriter.create<mlir::LLVM::FNegOp>(loc, eleTy, ip); | ||||||
| 3517 | auto r = rewriter.create<mlir::LLVM::InsertValueOp>(loc, o0, nrp, 0); | ||||||
| 3518 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>(neg, r, nip, 1); | ||||||
| 3519 | return mlir::success(); | ||||||
| 3520 | } | ||||||
| 3521 | }; | ||||||
| 3522 | |||||||
| 3523 | /// Conversion pattern for operation that must be dead. The information in these | ||||||
| 3524 | /// operations is used by other operation. At this point they should not have | ||||||
| 3525 | /// anymore uses. | ||||||
| 3526 | /// These operations are normally dead after the pre-codegen pass. | ||||||
| 3527 | template <typename FromOp> | ||||||
| 3528 | struct MustBeDeadConversion : public FIROpConversion<FromOp> { | ||||||
| 3529 | explicit MustBeDeadConversion(fir::LLVMTypeConverter &lowering, | ||||||
| 3530 | const fir::FIRToLLVMPassOptions &options) | ||||||
| 3531 | : FIROpConversion<FromOp>(lowering, options) {} | ||||||
| 3532 | using OpAdaptor = typename FromOp::Adaptor; | ||||||
| 3533 | |||||||
| 3534 | mlir::LogicalResult | ||||||
| 3535 | matchAndRewrite(FromOp op, OpAdaptor adaptor, | ||||||
| 3536 | mlir::ConversionPatternRewriter &rewriter) const final { | ||||||
| 3537 | if (!op->getUses().empty()) | ||||||
| 3538 | return rewriter.notifyMatchFailure(op, "op must be dead"); | ||||||
| 3539 | rewriter.eraseOp(op); | ||||||
| 3540 | return mlir::success(); | ||||||
| 3541 | } | ||||||
| 3542 | }; | ||||||
| 3543 | |||||||
| 3544 | struct UnrealizedConversionCastOpConversion | ||||||
| 3545 | : public FIROpConversion<mlir::UnrealizedConversionCastOp> { | ||||||
| 3546 | using FIROpConversion::FIROpConversion; | ||||||
| 3547 | |||||||
| 3548 | mlir::LogicalResult | ||||||
| 3549 | matchAndRewrite(mlir::UnrealizedConversionCastOp op, OpAdaptor adaptor, | ||||||
| 3550 | mlir::ConversionPatternRewriter &rewriter) const override { | ||||||
| 3551 | assert(op.getOutputs().getTypes().size() == 1 && "expect a single type")(static_cast <bool> (op.getOutputs().getTypes().size() == 1 && "expect a single type") ? void (0) : __assert_fail ("op.getOutputs().getTypes().size() == 1 && \"expect a single type\"" , "flang/lib/Optimizer/CodeGen/CodeGen.cpp", 3551, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3552 | mlir::Type convertedType = convertType(op.getOutputs().getTypes()[0]); | ||||||
| 3553 | if (convertedType == adaptor.getInputs().getTypes()[0]) { | ||||||
| 3554 | rewriter.replaceOp(op, adaptor.getInputs()); | ||||||
| 3555 | return mlir::success(); | ||||||
| 3556 | } | ||||||
| 3557 | |||||||
| 3558 | convertedType = adaptor.getInputs().getTypes()[0]; | ||||||
| 3559 | if (convertedType == op.getOutputs().getType()[0]) { | ||||||
| 3560 | rewriter.replaceOp(op, adaptor.getInputs()); | ||||||
| 3561 | return mlir::success(); | ||||||
| 3562 | } | ||||||
| 3563 | return mlir::failure(); | ||||||
| 3564 | } | ||||||
| 3565 | }; | ||||||
| 3566 | |||||||
| 3567 | struct ShapeOpConversion : public MustBeDeadConversion<fir::ShapeOp> { | ||||||
| 3568 | using MustBeDeadConversion::MustBeDeadConversion; | ||||||
| 3569 | }; | ||||||
| 3570 | |||||||
| 3571 | struct ShapeShiftOpConversion : public MustBeDeadConversion<fir::ShapeShiftOp> { | ||||||
| 3572 | using MustBeDeadConversion::MustBeDeadConversion; | ||||||
| 3573 | }; | ||||||
| 3574 | |||||||
| 3575 | struct ShiftOpConversion : public MustBeDeadConversion<fir::ShiftOp> { | ||||||
| 3576 | using MustBeDeadConversion::MustBeDeadConversion; | ||||||
| 3577 | }; | ||||||
| 3578 | |||||||
| 3579 | struct SliceOpConversion : public MustBeDeadConversion<fir::SliceOp> { | ||||||
| 3580 | using MustBeDeadConversion::MustBeDeadConversion; | ||||||
| 3581 | }; | ||||||
| 3582 | |||||||
| 3583 | } // namespace | ||||||
| 3584 | |||||||
| 3585 | namespace { | ||||||
| 3586 | class RenameMSVCLibmCallees | ||||||
| 3587 | : public mlir::OpRewritePattern<mlir::LLVM::CallOp> { | ||||||
| 3588 | public: | ||||||
| 3589 | using OpRewritePattern::OpRewritePattern; | ||||||
| 3590 | |||||||
| 3591 | mlir::LogicalResult | ||||||
| 3592 | matchAndRewrite(mlir::LLVM::CallOp op, | ||||||
| 3593 | mlir::PatternRewriter &rewriter) const override { | ||||||
| 3594 | rewriter.startRootUpdate(op); | ||||||
| 3595 | auto callee = op.getCallee(); | ||||||
| 3596 | if (callee) | ||||||
| 3597 | if (callee->equals("hypotf")) | ||||||
| 3598 | op.setCalleeAttr(mlir::SymbolRefAttr::get(op.getContext(), "_hypotf")); | ||||||
| 3599 | |||||||
| 3600 | rewriter.finalizeRootUpdate(op); | ||||||
| 3601 | return mlir::success(); | ||||||
| 3602 | } | ||||||
| 3603 | }; | ||||||
| 3604 | |||||||
| 3605 | class RenameMSVCLibmFuncs | ||||||
| 3606 | : public mlir::OpRewritePattern<mlir::LLVM::LLVMFuncOp> { | ||||||
| 3607 | public: | ||||||
| 3608 | using OpRewritePattern::OpRewritePattern; | ||||||
| 3609 | |||||||
| 3610 | mlir::LogicalResult | ||||||
| 3611 | matchAndRewrite(mlir::LLVM::LLVMFuncOp op, | ||||||
| 3612 | mlir::PatternRewriter &rewriter) const override { | ||||||
| 3613 | rewriter.startRootUpdate(op); | ||||||
| 3614 | if (op.getSymName().equals("hypotf")) | ||||||
| 3615 | op.setSymNameAttr(rewriter.getStringAttr("_hypotf")); | ||||||
| 3616 | rewriter.finalizeRootUpdate(op); | ||||||
| 3617 | return mlir::success(); | ||||||
| 3618 | } | ||||||
| 3619 | }; | ||||||
| 3620 | } // namespace | ||||||
| 3621 | |||||||
| 3622 | namespace { | ||||||
| 3623 | /// Convert FIR dialect to LLVM dialect | ||||||
| 3624 | /// | ||||||
| 3625 | /// This pass lowers all FIR dialect operations to LLVM IR dialect. An | ||||||
| 3626 | /// MLIR pass is used to lower residual Std dialect to LLVM IR dialect. | ||||||
| 3627 | class FIRToLLVMLowering | ||||||
| 3628 | : public fir::impl::FIRToLLVMLoweringBase<FIRToLLVMLowering> { | ||||||
| 3629 | public: | ||||||
| 3630 | FIRToLLVMLowering() = default; | ||||||
| 3631 | FIRToLLVMLowering(fir::FIRToLLVMPassOptions options) : options{options} {} | ||||||
| 3632 | mlir::ModuleOp getModule() { return getOperation(); } | ||||||
| 3633 | |||||||
| 3634 | void runOnOperation() override final { | ||||||
| 3635 | auto mod = getModule(); | ||||||
| 3636 | if (!forcedTargetTriple.empty()) | ||||||
| 3637 | fir::setTargetTriple(mod, forcedTargetTriple); | ||||||
| 3638 | |||||||
| 3639 | // Run dynamic pass pipeline for converting Math dialect | ||||||
| 3640 | // operations into other dialects (llvm, func, etc.). | ||||||
| 3641 | // Some conversions of Math operations cannot be done | ||||||
| 3642 | // by just using conversion patterns. This is true for | ||||||
| 3643 | // conversions that affect the ModuleOp, e.g. create new | ||||||
| 3644 | // function operations in it. We have to run such conversions | ||||||
| 3645 | // as passes here. | ||||||
| 3646 | mlir::OpPassManager mathConvertionPM("builtin.module"); | ||||||
| 3647 | |||||||
| 3648 | // Convert math::FPowI operations to inline implementation | ||||||
| 3649 | // only if the exponent's width is greater than 32, otherwise, | ||||||
| 3650 | // it will be lowered to LLVM intrinsic operation by a later conversion. | ||||||
| 3651 | mlir::ConvertMathToFuncsOptions mathToFuncsOptions{}; | ||||||
| 3652 | mathToFuncsOptions.minWidthOfFPowIExponent = 33; | ||||||
| 3653 | mathConvertionPM.addPass( | ||||||
| 3654 | mlir::createConvertMathToFuncs(mathToFuncsOptions)); | ||||||
| 3655 | mathConvertionPM.addPass(mlir::createConvertComplexToStandardPass()); | ||||||
| 3656 | // Convert Math dialect operations into LLVM dialect operations. | ||||||
| 3657 | // There is no way to prefer MathToLLVM patterns over MathToLibm | ||||||
| 3658 | // patterns (applied below), so we have to run MathToLLVM conversion here. | ||||||
| 3659 | mathConvertionPM.addNestedPass<mlir::func::FuncOp>( | ||||||
| 3660 | mlir::createConvertMathToLLVMPass()); | ||||||
| 3661 | if (mlir::failed(runPipeline(mathConvertionPM, mod))) | ||||||
| 3662 | return signalPassFailure(); | ||||||
| 3663 | |||||||
| 3664 | auto *context = getModule().getContext(); | ||||||
| 3665 | fir::LLVMTypeConverter typeConverter{getModule(), | ||||||
| 3666 | options.applyTBAA || applyTBAA}; | ||||||
| 3667 | mlir::RewritePatternSet pattern(context); | ||||||
| 3668 | pattern.insert< | ||||||
| 3669 | AbsentOpConversion, AddcOpConversion, AddrOfOpConversion, | ||||||
| 3670 | AllocaOpConversion, AllocMemOpConversion, BoxAddrOpConversion, | ||||||
| 3671 | BoxCharLenOpConversion, BoxDimsOpConversion, BoxEleSizeOpConversion, | ||||||
| 3672 | BoxIsAllocOpConversion, BoxIsArrayOpConversion, BoxIsPtrOpConversion, | ||||||
| 3673 | BoxProcHostOpConversion, BoxRankOpConversion, BoxTypeCodeOpConversion, | ||||||
| 3674 | BoxTypeDescOpConversion, CallOpConversion, CmpcOpConversion, | ||||||
| 3675 | ConstcOpConversion, ConvertOpConversion, CoordinateOpConversion, | ||||||
| 3676 | DispatchTableOpConversion, DTEntryOpConversion, DivcOpConversion, | ||||||
| 3677 | EmboxOpConversion, EmboxCharOpConversion, EmboxProcOpConversion, | ||||||
| 3678 | ExtractValueOpConversion, FieldIndexOpConversion, FirEndOpConversion, | ||||||
| 3679 | FreeMemOpConversion, GlobalLenOpConversion, GlobalOpConversion, | ||||||
| 3680 | HasValueOpConversion, InsertOnRangeOpConversion, | ||||||
| 3681 | InsertValueOpConversion, IsPresentOpConversion, | ||||||
| 3682 | LenParamIndexOpConversion, LoadOpConversion, MulcOpConversion, | ||||||
| 3683 | NegcOpConversion, NoReassocOpConversion, SelectCaseOpConversion, | ||||||
| 3684 | SelectOpConversion, SelectRankOpConversion, SelectTypeOpConversion, | ||||||
| 3685 | ShapeOpConversion, ShapeShiftOpConversion, ShiftOpConversion, | ||||||
| 3686 | SliceOpConversion, StoreOpConversion, StringLitOpConversion, | ||||||
| 3687 | SubcOpConversion, TypeDescOpConversion, UnboxCharOpConversion, | ||||||
| 3688 | UnboxProcOpConversion, UndefOpConversion, UnreachableOpConversion, | ||||||
| 3689 | UnrealizedConversionCastOpConversion, XArrayCoorOpConversion, | ||||||
| 3690 | XEmboxOpConversion, XReboxOpConversion, ZeroOpConversion>(typeConverter, | ||||||
| 3691 | options); | ||||||
| 3692 | mlir::populateFuncToLLVMConversionPatterns(typeConverter, pattern); | ||||||
| 3693 | mlir::populateOpenACCToLLVMConversionPatterns(typeConverter, pattern); | ||||||
| 3694 | mlir::populateOpenMPToLLVMConversionPatterns(typeConverter, pattern); | ||||||
| 3695 | mlir::arith::populateArithToLLVMConversionPatterns(typeConverter, pattern); | ||||||
| 3696 | mlir::cf::populateControlFlowToLLVMConversionPatterns(typeConverter, | ||||||
| 3697 | pattern); | ||||||
| 3698 | // Math operations that have not been converted yet must be converted | ||||||
| 3699 | // to Libm. | ||||||
| 3700 | mlir::populateMathToLibmConversionPatterns(pattern); | ||||||
| 3701 | mlir::populateComplexToLLVMConversionPatterns(typeConverter, pattern); | ||||||
| 3702 | mlir::ConversionTarget target{*context}; | ||||||
| 3703 | target.addLegalDialect<mlir::LLVM::LLVMDialect>(); | ||||||
| 3704 | // The OpenMP dialect is legal for Operations without regions, for those | ||||||
| 3705 | // which contains regions it is legal if the region contains only the | ||||||
| 3706 | // LLVM dialect. Add OpenMP dialect as a legal dialect for conversion and | ||||||
| 3707 | // legalize conversion of OpenMP operations without regions. | ||||||
| 3708 | mlir::configureOpenMPToLLVMConversionLegality(target, typeConverter); | ||||||
| 3709 | target.addLegalDialect<mlir::omp::OpenMPDialect>(); | ||||||
| 3710 | target.addLegalDialect<mlir::acc::OpenACCDialect>(); | ||||||
| 3711 | |||||||
| 3712 | // required NOPs for applying a full conversion | ||||||
| 3713 | target.addLegalOp<mlir::ModuleOp>(); | ||||||
| 3714 | |||||||
| 3715 | // If we're on Windows, we might need to rename some libm calls. | ||||||
| 3716 | bool isMSVC = fir::getTargetTriple(mod).isOSMSVCRT(); | ||||||
| 3717 | if (isMSVC) { | ||||||
| 3718 | pattern.insert<RenameMSVCLibmCallees, RenameMSVCLibmFuncs>(context); | ||||||
| 3719 | |||||||
| 3720 | target.addDynamicallyLegalOp<mlir::LLVM::CallOp>( | ||||||
| 3721 | [](mlir::LLVM::CallOp op) { | ||||||
| 3722 | auto callee = op.getCallee(); | ||||||
| 3723 | if (!callee) | ||||||
| 3724 | return true; | ||||||
| 3725 | return !callee->equals("hypotf"); | ||||||
| 3726 | }); | ||||||
| 3727 | target.addDynamicallyLegalOp<mlir::LLVM::LLVMFuncOp>( | ||||||
| 3728 | [](mlir::LLVM::LLVMFuncOp op) { | ||||||
| 3729 | return !op.getSymName().equals("hypotf"); | ||||||
| 3730 | }); | ||||||
| 3731 | } | ||||||
| 3732 | |||||||
| 3733 | // apply the patterns | ||||||
| 3734 | if (mlir::failed(mlir::applyFullConversion(getModule(), target, | ||||||
| 3735 | std::move(pattern)))) { | ||||||
| 3736 | signalPassFailure(); | ||||||
| 3737 | } | ||||||
| 3738 | } | ||||||
| 3739 | |||||||
| 3740 | private: | ||||||
| 3741 | fir::FIRToLLVMPassOptions options; | ||||||
| 3742 | }; | ||||||
| 3743 | |||||||
| 3744 | /// Lower from LLVM IR dialect to proper LLVM-IR and dump the module | ||||||
| 3745 | struct LLVMIRLoweringPass | ||||||
| 3746 | : public mlir::PassWrapper<LLVMIRLoweringPass, | ||||||
| 3747 | mlir::OperationPass<mlir::ModuleOp>> { | ||||||
| 3748 | MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(LLVMIRLoweringPass)static ::mlir::TypeID resolveTypeID() { static ::mlir::SelfOwningTypeID id; return id; } static_assert( ::mlir::detail::InlineTypeIDResolver ::has_resolve_typeid< LLVMIRLoweringPass>::value, "`MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID` must be placed in a " "public section of `" "LLVMIRLoweringPass" "`"); | ||||||
| 3749 | |||||||
| 3750 | LLVMIRLoweringPass(llvm::raw_ostream &output, fir::LLVMIRLoweringPrinter p) | ||||||
| 3751 | : output{output}, printer{p} {} | ||||||
| 3752 | |||||||
| 3753 | mlir::ModuleOp getModule() { return getOperation(); } | ||||||
| 3754 | |||||||
| 3755 | void runOnOperation() override final { | ||||||
| 3756 | auto *ctx = getModule().getContext(); | ||||||
| 3757 | auto optName = getModule().getName(); | ||||||
| 3758 | llvm::LLVMContext llvmCtx; | ||||||
| 3759 | if (auto llvmModule = mlir::translateModuleToLLVMIR( | ||||||
| 3760 | getModule(), llvmCtx, optName ? *optName : "FIRModule")) { | ||||||
| 3761 | printer(*llvmModule, output); | ||||||
| 3762 | return; | ||||||
| 3763 | } | ||||||
| 3764 | |||||||
| 3765 | mlir::emitError(mlir::UnknownLoc::get(ctx), "could not emit LLVM-IR\n"); | ||||||
| 3766 | signalPassFailure(); | ||||||
| 3767 | } | ||||||
| 3768 | |||||||
| 3769 | private: | ||||||
| 3770 | llvm::raw_ostream &output; | ||||||
| 3771 | fir::LLVMIRLoweringPrinter printer; | ||||||
| 3772 | }; | ||||||
| 3773 | |||||||
| 3774 | } // namespace | ||||||
| 3775 | |||||||
| 3776 | std::unique_ptr<mlir::Pass> fir::createFIRToLLVMPass() { | ||||||
| 3777 | return std::make_unique<FIRToLLVMLowering>(); | ||||||
| 3778 | } | ||||||
| 3779 | |||||||
| 3780 | std::unique_ptr<mlir::Pass> | ||||||
| 3781 | fir::createFIRToLLVMPass(fir::FIRToLLVMPassOptions options) { | ||||||
| 3782 | return std::make_unique<FIRToLLVMLowering>(options); | ||||||
| 3783 | } | ||||||
| 3784 | |||||||
| 3785 | std::unique_ptr<mlir::Pass> | ||||||
| 3786 | fir::createLLVMDialectToLLVMPass(llvm::raw_ostream &output, | ||||||
| 3787 | fir::LLVMIRLoweringPrinter printer) { | ||||||
| 3788 | return std::make_unique<LLVMIRLoweringPass>(output, printer); | ||||||
| 3789 | } |
| 1 | //===- PatternMatch.h - PatternMatcher classes -------==---------*- C++ -*-===// |
| 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 | #ifndef MLIR_IR_PATTERNMATCH_H |
| 10 | #define MLIR_IR_PATTERNMATCH_H |
| 11 | |
| 12 | #include "mlir/IR/Builders.h" |
| 13 | #include "mlir/IR/BuiltinOps.h" |
| 14 | #include "llvm/ADT/FunctionExtras.h" |
| 15 | #include "llvm/Support/TypeName.h" |
| 16 | #include <optional> |
| 17 | |
| 18 | namespace mlir { |
| 19 | |
| 20 | class PatternRewriter; |
| 21 | |
| 22 | //===----------------------------------------------------------------------===// |
| 23 | // PatternBenefit class |
| 24 | //===----------------------------------------------------------------------===// |
| 25 | |
| 26 | /// This class represents the benefit of a pattern match in a unitless scheme |
| 27 | /// that ranges from 0 (very little benefit) to 65K. The most common unit to |
| 28 | /// use here is the "number of operations matched" by the pattern. |
| 29 | /// |
| 30 | /// This also has a sentinel representation that can be used for patterns that |
| 31 | /// fail to match. |
| 32 | /// |
| 33 | class PatternBenefit { |
| 34 | enum { ImpossibleToMatchSentinel = 65535 }; |
| 35 | |
| 36 | public: |
| 37 | PatternBenefit() = default; |
| 38 | PatternBenefit(unsigned benefit); |
| 39 | PatternBenefit(const PatternBenefit &) = default; |
| 40 | PatternBenefit &operator=(const PatternBenefit &) = default; |
| 41 | |
| 42 | static PatternBenefit impossibleToMatch() { return PatternBenefit(); } |
| 43 | bool isImpossibleToMatch() const { return *this == impossibleToMatch(); } |
| 44 | |
| 45 | /// If the corresponding pattern can match, return its benefit. If the |
| 46 | // corresponding pattern isImpossibleToMatch() then this aborts. |
| 47 | unsigned short getBenefit() const; |
| 48 | |
| 49 | bool operator==(const PatternBenefit &rhs) const { |
| 50 | return representation == rhs.representation; |
| 51 | } |
| 52 | bool operator!=(const PatternBenefit &rhs) const { return !(*this == rhs); } |
| 53 | bool operator<(const PatternBenefit &rhs) const { |
| 54 | return representation < rhs.representation; |
| 55 | } |
| 56 | bool operator>(const PatternBenefit &rhs) const { return rhs < *this; } |
| 57 | bool operator<=(const PatternBenefit &rhs) const { return !(*this > rhs); } |
| 58 | bool operator>=(const PatternBenefit &rhs) const { return !(*this < rhs); } |
| 59 | |
| 60 | private: |
| 61 | unsigned short representation{ImpossibleToMatchSentinel}; |
| 62 | }; |
| 63 | |
| 64 | //===----------------------------------------------------------------------===// |
| 65 | // Pattern |
| 66 | //===----------------------------------------------------------------------===// |
| 67 | |
| 68 | /// This class contains all of the data related to a pattern, but does not |
| 69 | /// contain any methods or logic for the actual matching. This class is solely |
| 70 | /// used to interface with the metadata of a pattern, such as the benefit or |
| 71 | /// root operation. |
| 72 | class Pattern { |
| 73 | /// This enum represents the kind of value used to select the root operations |
| 74 | /// that match this pattern. |
| 75 | enum class RootKind { |
| 76 | /// The pattern root matches "any" operation. |
| 77 | Any, |
| 78 | /// The pattern root is matched using a concrete operation name. |
| 79 | OperationName, |
| 80 | /// The pattern root is matched using an interface ID. |
| 81 | InterfaceID, |
| 82 | /// The patter root is matched using a trait ID. |
| 83 | TraitID |
| 84 | }; |
| 85 | |
| 86 | public: |
| 87 | /// Return a list of operations that may be generated when rewriting an |
| 88 | /// operation instance with this pattern. |
| 89 | ArrayRef<OperationName> getGeneratedOps() const { return generatedOps; } |
| 90 | |
| 91 | /// Return the root node that this pattern matches. Patterns that can match |
| 92 | /// multiple root types return std::nullopt. |
| 93 | std::optional<OperationName> getRootKind() const { |
| 94 | if (rootKind == RootKind::OperationName) |
| 95 | return OperationName::getFromOpaquePointer(rootValue); |
| 96 | return std::nullopt; |
| 97 | } |
| 98 | |
| 99 | /// Return the interface ID used to match the root operation of this pattern. |
| 100 | /// If the pattern does not use an interface ID for deciding the root match, |
| 101 | /// this returns std::nullopt. |
| 102 | std::optional<TypeID> getRootInterfaceID() const { |
| 103 | if (rootKind == RootKind::InterfaceID) |
| 104 | return TypeID::getFromOpaquePointer(rootValue); |
| 105 | return std::nullopt; |
| 106 | } |
| 107 | |
| 108 | /// Return the trait ID used to match the root operation of this pattern. |
| 109 | /// If the pattern does not use a trait ID for deciding the root match, this |
| 110 | /// returns std::nullopt. |
| 111 | std::optional<TypeID> getRootTraitID() const { |
| 112 | if (rootKind == RootKind::TraitID) |
| 113 | return TypeID::getFromOpaquePointer(rootValue); |
| 114 | return std::nullopt; |
| 115 | } |
| 116 | |
| 117 | /// Return the benefit (the inverse of "cost") of matching this pattern. The |
| 118 | /// benefit of a Pattern is always static - rewrites that may have dynamic |
| 119 | /// benefit can be instantiated multiple times (different Pattern instances) |
| 120 | /// for each benefit that they may return, and be guarded by different match |
| 121 | /// condition predicates. |
| 122 | PatternBenefit getBenefit() const { return benefit; } |
| 123 | |
| 124 | /// Returns true if this pattern is known to result in recursive application, |
| 125 | /// i.e. this pattern may generate IR that also matches this pattern, but is |
| 126 | /// known to bound the recursion. This signals to a rewrite driver that it is |
| 127 | /// safe to apply this pattern recursively to generated IR. |
| 128 | bool hasBoundedRewriteRecursion() const { |
| 129 | return contextAndHasBoundedRecursion.getInt(); |
| 130 | } |
| 131 | |
| 132 | /// Return the MLIRContext used to create this pattern. |
| 133 | MLIRContext *getContext() const { |
| 134 | return contextAndHasBoundedRecursion.getPointer(); |
| 135 | } |
| 136 | |
| 137 | /// Return a readable name for this pattern. This name should only be used for |
| 138 | /// debugging purposes, and may be empty. |
| 139 | StringRef getDebugName() const { return debugName; } |
| 140 | |
| 141 | /// Set the human readable debug name used for this pattern. This name will |
| 142 | /// only be used for debugging purposes. |
| 143 | void setDebugName(StringRef name) { debugName = name; } |
| 144 | |
| 145 | /// Return the set of debug labels attached to this pattern. |
| 146 | ArrayRef<StringRef> getDebugLabels() const { return debugLabels; } |
| 147 | |
| 148 | /// Add the provided debug labels to this pattern. |
| 149 | void addDebugLabels(ArrayRef<StringRef> labels) { |
| 150 | debugLabels.append(labels.begin(), labels.end()); |
| 151 | } |
| 152 | void addDebugLabels(StringRef label) { debugLabels.push_back(label); } |
| 153 | |
| 154 | protected: |
| 155 | /// This class acts as a special tag that makes the desire to match "any" |
| 156 | /// operation type explicit. This helps to avoid unnecessary usages of this |
| 157 | /// feature, and ensures that the user is making a conscious decision. |
| 158 | struct MatchAnyOpTypeTag {}; |
| 159 | /// This class acts as a special tag that makes the desire to match any |
| 160 | /// operation that implements a given interface explicit. This helps to avoid |
| 161 | /// unnecessary usages of this feature, and ensures that the user is making a |
| 162 | /// conscious decision. |
| 163 | struct MatchInterfaceOpTypeTag {}; |
| 164 | /// This class acts as a special tag that makes the desire to match any |
| 165 | /// operation that implements a given trait explicit. This helps to avoid |
| 166 | /// unnecessary usages of this feature, and ensures that the user is making a |
| 167 | /// conscious decision. |
| 168 | struct MatchTraitOpTypeTag {}; |
| 169 | |
| 170 | /// Construct a pattern with a certain benefit that matches the operation |
| 171 | /// with the given root name. |
| 172 | Pattern(StringRef rootName, PatternBenefit benefit, MLIRContext *context, |
| 173 | ArrayRef<StringRef> generatedNames = {}); |
| 174 | /// Construct a pattern that may match any operation type. `generatedNames` |
| 175 | /// contains the names of operations that may be generated during a successful |
| 176 | /// rewrite. `MatchAnyOpTypeTag` is just a tag to ensure that the "match any" |
| 177 | /// behavior is what the user actually desired, `MatchAnyOpTypeTag()` should |
| 178 | /// always be supplied here. |
| 179 | Pattern(MatchAnyOpTypeTag tag, PatternBenefit benefit, MLIRContext *context, |
| 180 | ArrayRef<StringRef> generatedNames = {}); |
| 181 | /// Construct a pattern that may match any operation that implements the |
| 182 | /// interface defined by the provided `interfaceID`. `generatedNames` contains |
| 183 | /// the names of operations that may be generated during a successful rewrite. |
| 184 | /// `MatchInterfaceOpTypeTag` is just a tag to ensure that the "match |
| 185 | /// interface" behavior is what the user actually desired, |
| 186 | /// `MatchInterfaceOpTypeTag()` should always be supplied here. |
| 187 | Pattern(MatchInterfaceOpTypeTag tag, TypeID interfaceID, |
| 188 | PatternBenefit benefit, MLIRContext *context, |
| 189 | ArrayRef<StringRef> generatedNames = {}); |
| 190 | /// Construct a pattern that may match any operation that implements the |
| 191 | /// trait defined by the provided `traitID`. `generatedNames` contains the |
| 192 | /// names of operations that may be generated during a successful rewrite. |
| 193 | /// `MatchTraitOpTypeTag` is just a tag to ensure that the "match trait" |
| 194 | /// behavior is what the user actually desired, `MatchTraitOpTypeTag()` should |
| 195 | /// always be supplied here. |
| 196 | Pattern(MatchTraitOpTypeTag tag, TypeID traitID, PatternBenefit benefit, |
| 197 | MLIRContext *context, ArrayRef<StringRef> generatedNames = {}); |
| 198 | |
| 199 | /// Set the flag detailing if this pattern has bounded rewrite recursion or |
| 200 | /// not. |
| 201 | void setHasBoundedRewriteRecursion(bool hasBoundedRecursionArg = true) { |
| 202 | contextAndHasBoundedRecursion.setInt(hasBoundedRecursionArg); |
| 203 | } |
| 204 | |
| 205 | private: |
| 206 | Pattern(const void *rootValue, RootKind rootKind, |
| 207 | ArrayRef<StringRef> generatedNames, PatternBenefit benefit, |
| 208 | MLIRContext *context); |
| 209 | |
| 210 | /// The value used to match the root operation of the pattern. |
| 211 | const void *rootValue; |
| 212 | RootKind rootKind; |
| 213 | |
| 214 | /// The expected benefit of matching this pattern. |
| 215 | const PatternBenefit benefit; |
| 216 | |
| 217 | /// The context this pattern was created from, and a boolean flag indicating |
| 218 | /// whether this pattern has bounded recursion or not. |
| 219 | llvm::PointerIntPair<MLIRContext *, 1, bool> contextAndHasBoundedRecursion; |
| 220 | |
| 221 | /// A list of the potential operations that may be generated when rewriting |
| 222 | /// an op with this pattern. |
| 223 | SmallVector<OperationName, 2> generatedOps; |
| 224 | |
| 225 | /// A readable name for this pattern. May be empty. |
| 226 | StringRef debugName; |
| 227 | |
| 228 | /// The set of debug labels attached to this pattern. |
| 229 | SmallVector<StringRef, 0> debugLabels; |
| 230 | }; |
| 231 | |
| 232 | //===----------------------------------------------------------------------===// |
| 233 | // RewritePattern |
| 234 | //===----------------------------------------------------------------------===// |
| 235 | |
| 236 | /// RewritePattern is the common base class for all DAG to DAG replacements. |
| 237 | /// There are two possible usages of this class: |
| 238 | /// * Multi-step RewritePattern with "match" and "rewrite" |
| 239 | /// - By overloading the "match" and "rewrite" functions, the user can |
| 240 | /// separate the concerns of matching and rewriting. |
| 241 | /// * Single-step RewritePattern with "matchAndRewrite" |
| 242 | /// - By overloading the "matchAndRewrite" function, the user can perform |
| 243 | /// the rewrite in the same call as the match. |
| 244 | /// |
| 245 | class RewritePattern : public Pattern { |
| 246 | public: |
| 247 | virtual ~RewritePattern() = default; |
| 248 | |
| 249 | /// Rewrite the IR rooted at the specified operation with the result of |
| 250 | /// this pattern, generating any new operations with the specified |
| 251 | /// builder. If an unexpected error is encountered (an internal |
| 252 | /// compiler error), it is emitted through the normal MLIR diagnostic |
| 253 | /// hooks and the IR is left in a valid state. |
| 254 | virtual void rewrite(Operation *op, PatternRewriter &rewriter) const; |
| 255 | |
| 256 | /// Attempt to match against code rooted at the specified operation, |
| 257 | /// which is the same operation code as getRootKind(). |
| 258 | virtual LogicalResult match(Operation *op) const; |
| 259 | |
| 260 | /// Attempt to match against code rooted at the specified operation, |
| 261 | /// which is the same operation code as getRootKind(). If successful, this |
| 262 | /// function will automatically perform the rewrite. |
| 263 | virtual LogicalResult matchAndRewrite(Operation *op, |
| 264 | PatternRewriter &rewriter) const { |
| 265 | if (succeeded(match(op))) { |
| 266 | rewrite(op, rewriter); |
| 267 | return success(); |
| 268 | } |
| 269 | return failure(); |
| 270 | } |
| 271 | |
| 272 | /// This method provides a convenient interface for creating and initializing |
| 273 | /// derived rewrite patterns of the given type `T`. |
| 274 | template <typename T, typename... Args> |
| 275 | static std::unique_ptr<T> create(Args &&...args) { |
| 276 | std::unique_ptr<T> pattern = |
| 277 | std::make_unique<T>(std::forward<Args>(args)...); |
| 278 | initializePattern<T>(*pattern); |
| 279 | |
| 280 | // Set a default debug name if one wasn't provided. |
| 281 | if (pattern->getDebugName().empty()) |
| 282 | pattern->setDebugName(llvm::getTypeName<T>()); |
| 283 | return pattern; |
| 284 | } |
| 285 | |
| 286 | protected: |
| 287 | /// Inherit the base constructors from `Pattern`. |
| 288 | using Pattern::Pattern; |
| 289 | |
| 290 | private: |
| 291 | /// Trait to check if T provides a `getOperationName` method. |
| 292 | template <typename T, typename... Args> |
| 293 | using has_initialize = decltype(std::declval<T>().initialize()); |
| 294 | template <typename T> |
| 295 | using detect_has_initialize = llvm::is_detected<has_initialize, T>; |
| 296 | |
| 297 | /// Initialize the derived pattern by calling its `initialize` method. |
| 298 | template <typename T> |
| 299 | static std::enable_if_t<detect_has_initialize<T>::value> |
| 300 | initializePattern(T &pattern) { |
| 301 | pattern.initialize(); |
| 302 | } |
| 303 | /// Empty derived pattern initializer for patterns that do not have an |
| 304 | /// initialize method. |
| 305 | template <typename T> |
| 306 | static std::enable_if_t<!detect_has_initialize<T>::value> |
| 307 | initializePattern(T &) {} |
| 308 | |
| 309 | /// An anchor for the virtual table. |
| 310 | virtual void anchor(); |
| 311 | }; |
| 312 | |
| 313 | namespace detail { |
| 314 | /// OpOrInterfaceRewritePatternBase is a wrapper around RewritePattern that |
| 315 | /// allows for matching and rewriting against an instance of a derived operation |
| 316 | /// class or Interface. |
| 317 | template <typename SourceOp> |
| 318 | struct OpOrInterfaceRewritePatternBase : public RewritePattern { |
| 319 | using RewritePattern::RewritePattern; |
| 320 | |
| 321 | /// Wrappers around the RewritePattern methods that pass the derived op type. |
| 322 | void rewrite(Operation *op, PatternRewriter &rewriter) const final { |
| 323 | rewrite(cast<SourceOp>(op), rewriter); |
| 324 | } |
| 325 | LogicalResult match(Operation *op) const final { |
| 326 | return match(cast<SourceOp>(op)); |
| 327 | } |
| 328 | LogicalResult matchAndRewrite(Operation *op, |
| 329 | PatternRewriter &rewriter) const final { |
| 330 | return matchAndRewrite(cast<SourceOp>(op), rewriter); |
| 331 | } |
| 332 | |
| 333 | /// Rewrite and Match methods that operate on the SourceOp type. These must be |
| 334 | /// overridden by the derived pattern class. |
| 335 | virtual void rewrite(SourceOp op, PatternRewriter &rewriter) const { |
| 336 | llvm_unreachable("must override rewrite or matchAndRewrite")::llvm::llvm_unreachable_internal("must override rewrite or matchAndRewrite" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 336); |
| 337 | } |
| 338 | virtual LogicalResult match(SourceOp op) const { |
| 339 | llvm_unreachable("must override match or matchAndRewrite")::llvm::llvm_unreachable_internal("must override match or matchAndRewrite" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 339); |
| 340 | } |
| 341 | virtual LogicalResult matchAndRewrite(SourceOp op, |
| 342 | PatternRewriter &rewriter) const { |
| 343 | if (succeeded(match(op))) { |
| 344 | rewrite(op, rewriter); |
| 345 | return success(); |
| 346 | } |
| 347 | return failure(); |
| 348 | } |
| 349 | }; |
| 350 | } // namespace detail |
| 351 | |
| 352 | /// OpRewritePattern is a wrapper around RewritePattern that allows for |
| 353 | /// matching and rewriting against an instance of a derived operation class as |
| 354 | /// opposed to a raw Operation. |
| 355 | template <typename SourceOp> |
| 356 | struct OpRewritePattern |
| 357 | : public detail::OpOrInterfaceRewritePatternBase<SourceOp> { |
| 358 | /// Patterns must specify the root operation name they match against, and can |
| 359 | /// also specify the benefit of the pattern matching and a list of generated |
| 360 | /// ops. |
| 361 | OpRewritePattern(MLIRContext *context, PatternBenefit benefit = 1, |
| 362 | ArrayRef<StringRef> generatedNames = {}) |
| 363 | : detail::OpOrInterfaceRewritePatternBase<SourceOp>( |
| 364 | SourceOp::getOperationName(), benefit, context, generatedNames) {} |
| 365 | }; |
| 366 | |
| 367 | /// OpInterfaceRewritePattern is a wrapper around RewritePattern that allows for |
| 368 | /// matching and rewriting against an instance of an operation interface instead |
| 369 | /// of a raw Operation. |
| 370 | template <typename SourceOp> |
| 371 | struct OpInterfaceRewritePattern |
| 372 | : public detail::OpOrInterfaceRewritePatternBase<SourceOp> { |
| 373 | OpInterfaceRewritePattern(MLIRContext *context, PatternBenefit benefit = 1) |
| 374 | : detail::OpOrInterfaceRewritePatternBase<SourceOp>( |
| 375 | Pattern::MatchInterfaceOpTypeTag(), SourceOp::getInterfaceID(), |
| 376 | benefit, context) {} |
| 377 | }; |
| 378 | |
| 379 | /// OpTraitRewritePattern is a wrapper around RewritePattern that allows for |
| 380 | /// matching and rewriting against instances of an operation that possess a |
| 381 | /// given trait. |
| 382 | template <template <typename> class TraitType> |
| 383 | class OpTraitRewritePattern : public RewritePattern { |
| 384 | public: |
| 385 | OpTraitRewritePattern(MLIRContext *context, PatternBenefit benefit = 1) |
| 386 | : RewritePattern(Pattern::MatchTraitOpTypeTag(), TypeID::get<TraitType>(), |
| 387 | benefit, context) {} |
| 388 | }; |
| 389 | |
| 390 | //===----------------------------------------------------------------------===// |
| 391 | // RewriterBase |
| 392 | //===----------------------------------------------------------------------===// |
| 393 | |
| 394 | /// This class coordinates the application of a rewrite on a set of IR, |
| 395 | /// providing a way for clients to track mutations and create new operations. |
| 396 | /// This class serves as a common API for IR mutation between pattern rewrites |
| 397 | /// and non-pattern rewrites, and facilitates the development of shared |
| 398 | /// IR transformation utilities. |
| 399 | class RewriterBase : public OpBuilder { |
| 400 | public: |
| 401 | struct Listener : public OpBuilder::Listener { |
| 402 | Listener() |
| 403 | : OpBuilder::Listener(ListenerBase::Kind::RewriterBaseListener) {} |
| 404 | |
| 405 | /// Notify the listener that the specified operation was modified in-place. |
| 406 | virtual void notifyOperationModified(Operation *op) {} |
| 407 | |
| 408 | /// Notify the listener that the specified operation is about to be replaced |
| 409 | /// with the set of values potentially produced by new operations. This is |
| 410 | /// called before the uses of the operation have been changed. |
| 411 | virtual void notifyOperationReplaced(Operation *op, |
| 412 | ValueRange replacement) {} |
| 413 | |
| 414 | /// This is called on an operation that a rewrite is removing, right before |
| 415 | /// the operation is deleted. At this point, the operation has zero uses. |
| 416 | virtual void notifyOperationRemoved(Operation *op) {} |
| 417 | |
| 418 | /// Notify the listener that the pattern failed to match the given |
| 419 | /// operation, and provide a callback to populate a diagnostic with the |
| 420 | /// reason why the failure occurred. This method allows for derived |
| 421 | /// listeners to optionally hook into the reason why a rewrite failed, and |
| 422 | /// display it to users. |
| 423 | virtual LogicalResult |
| 424 | notifyMatchFailure(Location loc, |
| 425 | function_ref<void(Diagnostic &)> reasonCallback) { |
| 426 | return failure(); |
| 427 | } |
| 428 | |
| 429 | static bool classof(const OpBuilder::Listener *base); |
| 430 | }; |
| 431 | |
| 432 | /// Move the blocks that belong to "region" before the given position in |
| 433 | /// another region "parent". The two regions must be different. The caller |
| 434 | /// is responsible for creating or updating the operation transferring flow |
| 435 | /// of control to the region and passing it the correct block arguments. |
| 436 | virtual void inlineRegionBefore(Region ®ion, Region &parent, |
| 437 | Region::iterator before); |
| 438 | void inlineRegionBefore(Region ®ion, Block *before); |
| 439 | |
| 440 | /// Clone the blocks that belong to "region" before the given position in |
| 441 | /// another region "parent". The two regions must be different. The caller is |
| 442 | /// responsible for creating or updating the operation transferring flow of |
| 443 | /// control to the region and passing it the correct block arguments. |
| 444 | virtual void cloneRegionBefore(Region ®ion, Region &parent, |
| 445 | Region::iterator before, IRMapping &mapping); |
| 446 | void cloneRegionBefore(Region ®ion, Region &parent, |
| 447 | Region::iterator before); |
| 448 | void cloneRegionBefore(Region ®ion, Block *before); |
| 449 | |
| 450 | /// This method replaces the uses of the results of `op` with the values in |
| 451 | /// `newValues` when the provided `functor` returns true for a specific use. |
| 452 | /// The number of values in `newValues` is required to match the number of |
| 453 | /// results of `op`. `allUsesReplaced`, if non-null, is set to true if all of |
| 454 | /// the uses of `op` were replaced. Note that in some rewriters, the given |
| 455 | /// 'functor' may be stored beyond the lifetime of the rewrite being applied. |
| 456 | /// As such, the function should not capture by reference and instead use |
| 457 | /// value capture as necessary. |
| 458 | virtual void |
| 459 | replaceOpWithIf(Operation *op, ValueRange newValues, bool *allUsesReplaced, |
| 460 | llvm::unique_function<bool(OpOperand &) const> functor); |
| 461 | void replaceOpWithIf(Operation *op, ValueRange newValues, |
| 462 | llvm::unique_function<bool(OpOperand &) const> functor) { |
| 463 | replaceOpWithIf(op, newValues, /*allUsesReplaced=*/nullptr, |
| 464 | std::move(functor)); |
| 465 | } |
| 466 | |
| 467 | /// This method replaces the uses of the results of `op` with the values in |
| 468 | /// `newValues` when a use is nested within the given `block`. The number of |
| 469 | /// values in `newValues` is required to match the number of results of `op`. |
| 470 | /// If all uses of this operation are replaced, the operation is erased. |
| 471 | void replaceOpWithinBlock(Operation *op, ValueRange newValues, Block *block, |
| 472 | bool *allUsesReplaced = nullptr); |
| 473 | |
| 474 | /// This method replaces the results of the operation with the specified list |
| 475 | /// of values. The number of provided values must match the number of results |
| 476 | /// of the operation. |
| 477 | virtual void replaceOp(Operation *op, ValueRange newValues); |
| 478 | |
| 479 | /// Replaces the result op with a new op that is created without verification. |
| 480 | /// The result values of the two ops must be the same types. |
| 481 | template <typename OpTy, typename... Args> |
| 482 | OpTy replaceOpWithNewOp(Operation *op, Args &&...args) { |
| 483 | auto newOp = create<OpTy>(op->getLoc(), std::forward<Args>(args)...); |
| 484 | replaceOpWithResultsOfAnotherOp(op, newOp.getOperation()); |
| 485 | return newOp; |
| 486 | } |
| 487 | |
| 488 | /// This method erases an operation that is known to have no uses. |
| 489 | virtual void eraseOp(Operation *op); |
| 490 | |
| 491 | /// This method erases all operations in a block. |
| 492 | virtual void eraseBlock(Block *block); |
| 493 | |
| 494 | /// Inline the operations of block 'source' into block 'dest' before the given |
| 495 | /// position. The source block will be deleted and must have no uses. |
| 496 | /// 'argValues' is used to replace the block arguments of 'source'. |
| 497 | /// |
| 498 | /// If the source block is inserted at the end of the dest block, the dest |
| 499 | /// block must have no successors. Similarly, if the source block is inserted |
| 500 | /// somewhere in the middle (or beginning) of the dest block, the source block |
| 501 | /// must have no successors. Otherwise, the resulting IR would have |
| 502 | /// unreachable operations. |
| 503 | virtual void inlineBlockBefore(Block *source, Block *dest, |
| 504 | Block::iterator before, |
| 505 | ValueRange argValues = std::nullopt); |
| 506 | |
| 507 | /// Inline the operations of block 'source' before the operation 'op'. The |
| 508 | /// source block will be deleted and must have no uses. 'argValues' is used to |
| 509 | /// replace the block arguments of 'source' |
| 510 | /// |
| 511 | /// The source block must have no successors. Otherwise, the resulting IR |
| 512 | /// would have unreachable operations. |
| 513 | void inlineBlockBefore(Block *source, Operation *op, |
| 514 | ValueRange argValues = std::nullopt); |
| 515 | |
| 516 | /// Inline the operations of block 'source' into the end of block 'dest'. The |
| 517 | /// source block will be deleted and must have no uses. 'argValues' is used to |
| 518 | /// replace the block arguments of 'source' |
| 519 | /// |
| 520 | /// The dest block must have no successors. Otherwise, the resulting IR would |
| 521 | /// have unreachable operation. |
| 522 | void mergeBlocks(Block *source, Block *dest, |
| 523 | ValueRange argValues = std::nullopt); |
| 524 | |
| 525 | /// Split the operations starting at "before" (inclusive) out of the given |
| 526 | /// block into a new block, and return it. |
| 527 | virtual Block *splitBlock(Block *block, Block::iterator before); |
| 528 | |
| 529 | /// This method is used to notify the rewriter that an in-place operation |
| 530 | /// modification is about to happen. A call to this function *must* be |
| 531 | /// followed by a call to either `finalizeRootUpdate` or `cancelRootUpdate`. |
| 532 | /// This is a minor efficiency win (it avoids creating a new operation and |
| 533 | /// removing the old one) but also often allows simpler code in the client. |
| 534 | virtual void startRootUpdate(Operation *op) {} |
| 535 | |
| 536 | /// This method is used to signal the end of a root update on the given |
| 537 | /// operation. This can only be called on operations that were provided to a |
| 538 | /// call to `startRootUpdate`. |
| 539 | virtual void finalizeRootUpdate(Operation *op); |
| 540 | |
| 541 | /// This method cancels a pending root update. This can only be called on |
| 542 | /// operations that were provided to a call to `startRootUpdate`. |
| 543 | virtual void cancelRootUpdate(Operation *op) {} |
| 544 | |
| 545 | /// This method is a utility wrapper around a root update of an operation. It |
| 546 | /// wraps calls to `startRootUpdate` and `finalizeRootUpdate` around the given |
| 547 | /// callable. |
| 548 | template <typename CallableT> |
| 549 | void updateRootInPlace(Operation *root, CallableT &&callable) { |
| 550 | startRootUpdate(root); |
| 551 | callable(); |
| 552 | finalizeRootUpdate(root); |
| 553 | } |
| 554 | |
| 555 | /// Find uses of `from` and replace them with `to`. It also marks every |
| 556 | /// modified uses and notifies the rewriter that an in-place operation |
| 557 | /// modification is about to happen. |
| 558 | void replaceAllUsesWith(Value from, Value to) { |
| 559 | return replaceAllUsesWith(from.getImpl(), to); |
| 560 | } |
| 561 | template <typename OperandType, typename ValueT> |
| 562 | void replaceAllUsesWith(IRObjectWithUseList<OperandType> *from, ValueT &&to) { |
| 563 | for (OperandType &operand : llvm::make_early_inc_range(from->getUses())) { |
| 564 | Operation *op = operand.getOwner(); |
| 565 | updateRootInPlace(op, [&]() { operand.set(to); }); |
| 566 | } |
| 567 | } |
| 568 | void replaceAllUsesWith(ValueRange from, ValueRange to) { |
| 569 | assert(from.size() == to.size() && "incorrect number of replacements")(static_cast <bool> (from.size() == to.size() && "incorrect number of replacements") ? void (0) : __assert_fail ("from.size() == to.size() && \"incorrect number of replacements\"" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 569, __extension__ __PRETTY_FUNCTION__)); |
| 570 | for (auto it : llvm::zip(from, to)) |
| 571 | replaceAllUsesWith(std::get<0>(it), std::get<1>(it)); |
| 572 | } |
| 573 | |
| 574 | /// Find uses of `from` and replace them with `to` if the `functor` returns |
| 575 | /// true. It also marks every modified uses and notifies the rewriter that an |
| 576 | /// in-place operation modification is about to happen. |
| 577 | void replaceUsesWithIf(Value from, Value to, |
| 578 | function_ref<bool(OpOperand &)> functor); |
| 579 | |
| 580 | /// Find uses of `from` and replace them with `to` except if the user is |
| 581 | /// `exceptedUser`. It also marks every modified uses and notifies the |
| 582 | /// rewriter that an in-place operation modification is about to happen. |
| 583 | void replaceAllUsesExcept(Value from, Value to, Operation *exceptedUser) { |
| 584 | return replaceUsesWithIf(from, to, [&](OpOperand &use) { |
| 585 | Operation *user = use.getOwner(); |
| 586 | return user != exceptedUser; |
| 587 | }); |
| 588 | } |
| 589 | |
| 590 | /// Used to notify the rewriter that the IR failed to be rewritten because of |
| 591 | /// a match failure, and provide a callback to populate a diagnostic with the |
| 592 | /// reason why the failure occurred. This method allows for derived rewriters |
| 593 | /// to optionally hook into the reason why a rewrite failed, and display it to |
| 594 | /// users. |
| 595 | template <typename CallbackT> |
| 596 | std::enable_if_t<!std::is_convertible<CallbackT, Twine>::value, LogicalResult> |
| 597 | notifyMatchFailure(Location loc, CallbackT &&reasonCallback) { |
| 598 | #ifndef NDEBUG |
| 599 | if (auto *rewriteListener = dyn_cast_if_present<Listener>(listener)) |
| 600 | return rewriteListener->notifyMatchFailure( |
| 601 | loc, function_ref<void(Diagnostic &)>(reasonCallback)); |
| 602 | return failure(); |
| 603 | #else |
| 604 | return failure(); |
| 605 | #endif |
| 606 | } |
| 607 | template <typename CallbackT> |
| 608 | std::enable_if_t<!std::is_convertible<CallbackT, Twine>::value, LogicalResult> |
| 609 | notifyMatchFailure(Operation *op, CallbackT &&reasonCallback) { |
| 610 | if (auto *rewriteListener = dyn_cast_if_present<Listener>(listener)) |
| 611 | return rewriteListener->notifyMatchFailure( |
| 612 | op->getLoc(), function_ref<void(Diagnostic &)>(reasonCallback)); |
| 613 | return failure(); |
| 614 | } |
| 615 | template <typename ArgT> |
| 616 | LogicalResult notifyMatchFailure(ArgT &&arg, const Twine &msg) { |
| 617 | return notifyMatchFailure(std::forward<ArgT>(arg), |
| 618 | [&](Diagnostic &diag) { diag << msg; }); |
| 619 | } |
| 620 | template <typename ArgT> |
| 621 | LogicalResult notifyMatchFailure(ArgT &&arg, const char *msg) { |
| 622 | return notifyMatchFailure(std::forward<ArgT>(arg), Twine(msg)); |
| 623 | } |
| 624 | |
| 625 | protected: |
| 626 | /// Initialize the builder. |
| 627 | explicit RewriterBase(MLIRContext *ctx, |
| 628 | OpBuilder::Listener *listener = nullptr) |
| 629 | : OpBuilder(ctx, listener) {} |
| 630 | explicit RewriterBase(const OpBuilder &otherBuilder) |
| 631 | : OpBuilder(otherBuilder) {} |
| 632 | virtual ~RewriterBase(); |
| 633 | |
| 634 | private: |
| 635 | void operator=(const RewriterBase &) = delete; |
| 636 | RewriterBase(const RewriterBase &) = delete; |
| 637 | |
| 638 | /// 'op' and 'newOp' are known to have the same number of results, replace the |
| 639 | /// uses of op with uses of newOp. |
| 640 | void replaceOpWithResultsOfAnotherOp(Operation *op, Operation *newOp); |
| 641 | }; |
| 642 | |
| 643 | //===----------------------------------------------------------------------===// |
| 644 | // IRRewriter |
| 645 | //===----------------------------------------------------------------------===// |
| 646 | |
| 647 | /// This class coordinates rewriting a piece of IR outside of a pattern rewrite, |
| 648 | /// providing a way to keep track of the mutations made to the IR. This class |
| 649 | /// should only be used in situations where another `RewriterBase` instance, |
| 650 | /// such as a `PatternRewriter`, is not available. |
| 651 | class IRRewriter : public RewriterBase { |
| 652 | public: |
| 653 | explicit IRRewriter(MLIRContext *ctx, OpBuilder::Listener *listener = nullptr) |
| 654 | : RewriterBase(ctx, listener) {} |
| 655 | explicit IRRewriter(const OpBuilder &builder) : RewriterBase(builder) {} |
| 656 | }; |
| 657 | |
| 658 | //===----------------------------------------------------------------------===// |
| 659 | // PatternRewriter |
| 660 | //===----------------------------------------------------------------------===// |
| 661 | |
| 662 | /// A special type of `RewriterBase` that coordinates the application of a |
| 663 | /// rewrite pattern on the current IR being matched, providing a way to keep |
| 664 | /// track of any mutations made. This class should be used to perform all |
| 665 | /// necessary IR mutations within a rewrite pattern, as the pattern driver may |
| 666 | /// be tracking various state that would be invalidated when a mutation takes |
| 667 | /// place. |
| 668 | class PatternRewriter : public RewriterBase { |
| 669 | public: |
| 670 | using RewriterBase::RewriterBase; |
| 671 | |
| 672 | /// A hook used to indicate if the pattern rewriter can recover from failure |
| 673 | /// during the rewrite stage of a pattern. For example, if the pattern |
| 674 | /// rewriter supports rollback, it may progress smoothly even if IR was |
| 675 | /// changed during the rewrite. |
| 676 | virtual bool canRecoverFromRewriteFailure() const { return false; } |
| 677 | }; |
| 678 | |
| 679 | //===----------------------------------------------------------------------===// |
| 680 | // PDL Patterns |
| 681 | //===----------------------------------------------------------------------===// |
| 682 | |
| 683 | //===----------------------------------------------------------------------===// |
| 684 | // PDLValue |
| 685 | |
| 686 | /// Storage type of byte-code interpreter values. These are passed to constraint |
| 687 | /// functions as arguments. |
| 688 | class PDLValue { |
| 689 | public: |
| 690 | /// The underlying kind of a PDL value. |
| 691 | enum class Kind { Attribute, Operation, Type, TypeRange, Value, ValueRange }; |
| 692 | |
| 693 | /// Construct a new PDL value. |
| 694 | PDLValue(const PDLValue &other) = default; |
| 695 | PDLValue(std::nullptr_t = nullptr) {} |
| 696 | PDLValue(Attribute value) |
| 697 | : value(value.getAsOpaquePointer()), kind(Kind::Attribute) {} |
| 698 | PDLValue(Operation *value) : value(value), kind(Kind::Operation) {} |
| 699 | PDLValue(Type value) : value(value.getAsOpaquePointer()), kind(Kind::Type) {} |
| 700 | PDLValue(TypeRange *value) : value(value), kind(Kind::TypeRange) {} |
| 701 | PDLValue(Value value) |
| 702 | : value(value.getAsOpaquePointer()), kind(Kind::Value) {} |
| 703 | PDLValue(ValueRange *value) : value(value), kind(Kind::ValueRange) {} |
| 704 | |
| 705 | /// Returns true if the type of the held value is `T`. |
| 706 | template <typename T> |
| 707 | bool isa() const { |
| 708 | assert(value && "isa<> used on a null value")(static_cast <bool> (value && "isa<> used on a null value" ) ? void (0) : __assert_fail ("value && \"isa<> used on a null value\"" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 708, __extension__ __PRETTY_FUNCTION__)); |
| 709 | return kind == getKindOf<T>(); |
| 710 | } |
| 711 | |
| 712 | /// Attempt to dynamically cast this value to type `T`, returns null if this |
| 713 | /// value is not an instance of `T`. |
| 714 | template <typename T, |
| 715 | typename ResultT = std::conditional_t< |
| 716 | std::is_convertible<T, bool>::value, T, std::optional<T>>> |
| 717 | ResultT dyn_cast() const { |
| 718 | return isa<T>() ? castImpl<T>() : ResultT(); |
| 719 | } |
| 720 | |
| 721 | /// Cast this value to type `T`, asserts if this value is not an instance of |
| 722 | /// `T`. |
| 723 | template <typename T> |
| 724 | T cast() const { |
| 725 | assert(isa<T>() && "expected value to be of type `T`")(static_cast <bool> (isa<T>() && "expected value to be of type `T`" ) ? void (0) : __assert_fail ("isa<T>() && \"expected value to be of type `T`\"" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 725, __extension__ __PRETTY_FUNCTION__)); |
| 726 | return castImpl<T>(); |
| 727 | } |
| 728 | |
| 729 | /// Get an opaque pointer to the value. |
| 730 | const void *getAsOpaquePointer() const { return value; } |
| 731 | |
| 732 | /// Return if this value is null or not. |
| 733 | explicit operator bool() const { return value; } |
| 734 | |
| 735 | /// Return the kind of this value. |
| 736 | Kind getKind() const { return kind; } |
| 737 | |
| 738 | /// Print this value to the provided output stream. |
| 739 | void print(raw_ostream &os) const; |
| 740 | |
| 741 | /// Print the specified value kind to an output stream. |
| 742 | static void print(raw_ostream &os, Kind kind); |
| 743 | |
| 744 | private: |
| 745 | /// Find the index of a given type in a range of other types. |
| 746 | template <typename...> |
| 747 | struct index_of_t; |
| 748 | template <typename T, typename... R> |
| 749 | struct index_of_t<T, T, R...> : std::integral_constant<size_t, 0> {}; |
| 750 | template <typename T, typename F, typename... R> |
| 751 | struct index_of_t<T, F, R...> |
| 752 | : std::integral_constant<size_t, 1 + index_of_t<T, R...>::value> {}; |
| 753 | |
| 754 | /// Return the kind used for the given T. |
| 755 | template <typename T> |
| 756 | static Kind getKindOf() { |
| 757 | return static_cast<Kind>(index_of_t<T, Attribute, Operation *, Type, |
| 758 | TypeRange, Value, ValueRange>::value); |
| 759 | } |
| 760 | |
| 761 | /// The internal implementation of `cast`, that returns the underlying value |
| 762 | /// as the given type `T`. |
| 763 | template <typename T> |
| 764 | std::enable_if_t<llvm::is_one_of<T, Attribute, Type, Value>::value, T> |
| 765 | castImpl() const { |
| 766 | return T::getFromOpaquePointer(value); |
| 767 | } |
| 768 | template <typename T> |
| 769 | std::enable_if_t<llvm::is_one_of<T, TypeRange, ValueRange>::value, T> |
| 770 | castImpl() const { |
| 771 | return *reinterpret_cast<T *>(const_cast<void *>(value)); |
| 772 | } |
| 773 | template <typename T> |
| 774 | std::enable_if_t<std::is_pointer<T>::value, T> castImpl() const { |
| 775 | return reinterpret_cast<T>(const_cast<void *>(value)); |
| 776 | } |
| 777 | |
| 778 | /// The internal opaque representation of a PDLValue. |
| 779 | const void *value{nullptr}; |
| 780 | /// The kind of the opaque value. |
| 781 | Kind kind{Kind::Attribute}; |
| 782 | }; |
| 783 | |
| 784 | inline raw_ostream &operator<<(raw_ostream &os, PDLValue value) { |
| 785 | value.print(os); |
| 786 | return os; |
| 787 | } |
| 788 | |
| 789 | inline raw_ostream &operator<<(raw_ostream &os, PDLValue::Kind kind) { |
| 790 | PDLValue::print(os, kind); |
| 791 | return os; |
| 792 | } |
| 793 | |
| 794 | //===----------------------------------------------------------------------===// |
| 795 | // PDLResultList |
| 796 | |
| 797 | /// The class represents a list of PDL results, returned by a native rewrite |
| 798 | /// method. It provides the mechanism with which to pass PDLValues back to the |
| 799 | /// PDL bytecode. |
| 800 | class PDLResultList { |
| 801 | public: |
| 802 | /// Push a new Attribute value onto the result list. |
| 803 | void push_back(Attribute value) { results.push_back(value); } |
| 804 | |
| 805 | /// Push a new Operation onto the result list. |
| 806 | void push_back(Operation *value) { results.push_back(value); } |
| 807 | |
| 808 | /// Push a new Type onto the result list. |
| 809 | void push_back(Type value) { results.push_back(value); } |
| 810 | |
| 811 | /// Push a new TypeRange onto the result list. |
| 812 | void push_back(TypeRange value) { |
| 813 | // The lifetime of a TypeRange can't be guaranteed, so we'll need to |
| 814 | // allocate a storage for it. |
| 815 | llvm::OwningArrayRef<Type> storage(value.size()); |
| 816 | llvm::copy(value, storage.begin()); |
| 817 | allocatedTypeRanges.emplace_back(std::move(storage)); |
| 818 | typeRanges.push_back(allocatedTypeRanges.back()); |
| 819 | results.push_back(&typeRanges.back()); |
| 820 | } |
| 821 | void push_back(ValueTypeRange<OperandRange> value) { |
| 822 | typeRanges.push_back(value); |
| 823 | results.push_back(&typeRanges.back()); |
| 824 | } |
| 825 | void push_back(ValueTypeRange<ResultRange> value) { |
| 826 | typeRanges.push_back(value); |
| 827 | results.push_back(&typeRanges.back()); |
| 828 | } |
| 829 | |
| 830 | /// Push a new Value onto the result list. |
| 831 | void push_back(Value value) { results.push_back(value); } |
| 832 | |
| 833 | /// Push a new ValueRange onto the result list. |
| 834 | void push_back(ValueRange value) { |
| 835 | // The lifetime of a ValueRange can't be guaranteed, so we'll need to |
| 836 | // allocate a storage for it. |
| 837 | llvm::OwningArrayRef<Value> storage(value.size()); |
| 838 | llvm::copy(value, storage.begin()); |
| 839 | allocatedValueRanges.emplace_back(std::move(storage)); |
| 840 | valueRanges.push_back(allocatedValueRanges.back()); |
| 841 | results.push_back(&valueRanges.back()); |
| 842 | } |
| 843 | void push_back(OperandRange value) { |
| 844 | valueRanges.push_back(value); |
| 845 | results.push_back(&valueRanges.back()); |
| 846 | } |
| 847 | void push_back(ResultRange value) { |
| 848 | valueRanges.push_back(value); |
| 849 | results.push_back(&valueRanges.back()); |
| 850 | } |
| 851 | |
| 852 | protected: |
| 853 | /// Create a new result list with the expected number of results. |
| 854 | PDLResultList(unsigned maxNumResults) { |
| 855 | // For now just reserve enough space for all of the results. We could do |
| 856 | // separate counts per range type, but it isn't really worth it unless there |
| 857 | // are a "large" number of results. |
| 858 | typeRanges.reserve(maxNumResults); |
| 859 | valueRanges.reserve(maxNumResults); |
| 860 | } |
| 861 | |
| 862 | /// The PDL results held by this list. |
| 863 | SmallVector<PDLValue> results; |
| 864 | /// Memory used to store ranges held by the list. |
| 865 | SmallVector<TypeRange> typeRanges; |
| 866 | SmallVector<ValueRange> valueRanges; |
| 867 | /// Memory allocated to store ranges in the result list whose lifetime was |
| 868 | /// generated in the native function. |
| 869 | SmallVector<llvm::OwningArrayRef<Type>> allocatedTypeRanges; |
| 870 | SmallVector<llvm::OwningArrayRef<Value>> allocatedValueRanges; |
| 871 | }; |
| 872 | |
| 873 | //===----------------------------------------------------------------------===// |
| 874 | // PDLPatternConfig |
| 875 | |
| 876 | /// An individual configuration for a pattern, which can be accessed by native |
| 877 | /// functions via the PDLPatternConfigSet. This allows for injecting additional |
| 878 | /// configuration into PDL patterns that is specific to certain compilation |
| 879 | /// flows. |
| 880 | class PDLPatternConfig { |
| 881 | public: |
| 882 | virtual ~PDLPatternConfig() = default; |
| 883 | |
| 884 | /// Hooks that are invoked at the beginning and end of a rewrite of a matched |
| 885 | /// pattern. These can be used to setup any specific state necessary for the |
| 886 | /// rewrite. |
| 887 | virtual void notifyRewriteBegin(PatternRewriter &rewriter) {} |
| 888 | virtual void notifyRewriteEnd(PatternRewriter &rewriter) {} |
| 889 | |
| 890 | /// Return the TypeID that represents this configuration. |
| 891 | TypeID getTypeID() const { return id; } |
| 892 | |
| 893 | protected: |
| 894 | PDLPatternConfig(TypeID id) : id(id) {} |
| 895 | |
| 896 | private: |
| 897 | TypeID id; |
| 898 | }; |
| 899 | |
| 900 | /// This class provides a base class for users implementing a type of pattern |
| 901 | /// configuration. |
| 902 | template <typename T> |
| 903 | class PDLPatternConfigBase : public PDLPatternConfig { |
| 904 | public: |
| 905 | /// Support LLVM style casting. |
| 906 | static bool classof(const PDLPatternConfig *config) { |
| 907 | return config->getTypeID() == getConfigID(); |
| 908 | } |
| 909 | |
| 910 | /// Return the type id used for this configuration. |
| 911 | static TypeID getConfigID() { return TypeID::get<T>(); } |
| 912 | |
| 913 | protected: |
| 914 | PDLPatternConfigBase() : PDLPatternConfig(getConfigID()) {} |
| 915 | }; |
| 916 | |
| 917 | /// This class contains a set of configurations for a specific pattern. |
| 918 | /// Configurations are uniqued by TypeID, meaning that only one configuration of |
| 919 | /// each type is allowed. |
| 920 | class PDLPatternConfigSet { |
| 921 | public: |
| 922 | PDLPatternConfigSet() = default; |
| 923 | |
| 924 | /// Construct a set with the given configurations. |
| 925 | template <typename... ConfigsT> |
| 926 | PDLPatternConfigSet(ConfigsT &&...configs) { |
| 927 | (addConfig(std::forward<ConfigsT>(configs)), ...); |
| 928 | } |
| 929 | |
| 930 | /// Get the configuration defined by the given type. Asserts that the |
| 931 | /// configuration of the provided type exists. |
| 932 | template <typename T> |
| 933 | const T &get() const { |
| 934 | const T *config = tryGet<T>(); |
| 935 | assert(config && "configuration not found")(static_cast <bool> (config && "configuration not found" ) ? void (0) : __assert_fail ("config && \"configuration not found\"" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 935, __extension__ __PRETTY_FUNCTION__)); |
| 936 | return *config; |
| 937 | } |
| 938 | |
| 939 | /// Get the configuration defined by the given type, returns nullptr if the |
| 940 | /// configuration does not exist. |
| 941 | template <typename T> |
| 942 | const T *tryGet() const { |
| 943 | for (const auto &configIt : configs) |
| 944 | if (const T *config = dyn_cast<T>(configIt.get())) |
| 945 | return config; |
| 946 | return nullptr; |
| 947 | } |
| 948 | |
| 949 | /// Notify the configurations within this set at the beginning or end of a |
| 950 | /// rewrite of a matched pattern. |
| 951 | void notifyRewriteBegin(PatternRewriter &rewriter) { |
| 952 | for (const auto &config : configs) |
| 953 | config->notifyRewriteBegin(rewriter); |
| 954 | } |
| 955 | void notifyRewriteEnd(PatternRewriter &rewriter) { |
| 956 | for (const auto &config : configs) |
| 957 | config->notifyRewriteEnd(rewriter); |
| 958 | } |
| 959 | |
| 960 | protected: |
| 961 | /// Add a configuration to the set. |
| 962 | template <typename T> |
| 963 | void addConfig(T &&config) { |
| 964 | assert(!tryGet<std::decay_t<T>>() && "configuration already exists")(static_cast <bool> (!tryGet<std::decay_t<T>> () && "configuration already exists") ? void (0) : __assert_fail ("!tryGet<std::decay_t<T>>() && \"configuration already exists\"" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 964, __extension__ __PRETTY_FUNCTION__)); |
| 965 | configs.emplace_back( |
| 966 | std::make_unique<std::decay_t<T>>(std::forward<T>(config))); |
| 967 | } |
| 968 | |
| 969 | /// The set of configurations for this pattern. This uses a vector instead of |
| 970 | /// a map with the expectation that the number of configurations per set is |
| 971 | /// small (<= 1). |
| 972 | SmallVector<std::unique_ptr<PDLPatternConfig>> configs; |
| 973 | }; |
| 974 | |
| 975 | //===----------------------------------------------------------------------===// |
| 976 | // PDLPatternModule |
| 977 | |
| 978 | /// A generic PDL pattern constraint function. This function applies a |
| 979 | /// constraint to a given set of opaque PDLValue entities. Returns success if |
| 980 | /// the constraint successfully held, failure otherwise. |
| 981 | using PDLConstraintFunction = |
| 982 | std::function<LogicalResult(PatternRewriter &, ArrayRef<PDLValue>)>; |
| 983 | /// A native PDL rewrite function. This function performs a rewrite on the |
| 984 | /// given set of values. Any results from this rewrite that should be passed |
| 985 | /// back to PDL should be added to the provided result list. This method is only |
| 986 | /// invoked when the corresponding match was successful. Returns failure if an |
| 987 | /// invariant of the rewrite was broken (certain rewriters may recover from |
| 988 | /// partial pattern application). |
| 989 | using PDLRewriteFunction = std::function<LogicalResult( |
| 990 | PatternRewriter &, PDLResultList &, ArrayRef<PDLValue>)>; |
| 991 | |
| 992 | namespace detail { |
| 993 | namespace pdl_function_builder { |
| 994 | /// A utility variable that always resolves to false. This is useful for static |
| 995 | /// asserts that are always false, but only should fire in certain templated |
| 996 | /// constructs. For example, if a templated function should never be called, the |
| 997 | /// function could be defined as: |
| 998 | /// |
| 999 | /// template <typename T> |
| 1000 | /// void foo() { |
| 1001 | /// static_assert(always_false<T>, "This function should never be called"); |
| 1002 | /// } |
| 1003 | /// |
| 1004 | template <class... T> |
| 1005 | constexpr bool always_false = false; |
| 1006 | |
| 1007 | //===----------------------------------------------------------------------===// |
| 1008 | // PDL Function Builder: Type Processing |
| 1009 | //===----------------------------------------------------------------------===// |
| 1010 | |
| 1011 | /// This struct provides a convenient way to determine how to process a given |
| 1012 | /// type as either a PDL parameter, or a result value. This allows for |
| 1013 | /// supporting complex types in constraint and rewrite functions, without |
| 1014 | /// requiring the user to hand-write the necessary glue code themselves. |
| 1015 | /// Specializations of this class should implement the following methods to |
| 1016 | /// enable support as a PDL argument or result type: |
| 1017 | /// |
| 1018 | /// static LogicalResult verifyAsArg( |
| 1019 | /// function_ref<LogicalResult(const Twine &)> errorFn, PDLValue pdlValue, |
| 1020 | /// size_t argIdx); |
| 1021 | /// |
| 1022 | /// * This method verifies that the given PDLValue is valid for use as a |
| 1023 | /// value of `T`. |
| 1024 | /// |
| 1025 | /// static T processAsArg(PDLValue pdlValue); |
| 1026 | /// |
| 1027 | /// * This method processes the given PDLValue as a value of `T`. |
| 1028 | /// |
| 1029 | /// static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1030 | /// const T &value); |
| 1031 | /// |
| 1032 | /// * This method processes the given value of `T` as the result of a |
| 1033 | /// function invocation. The method should package the value into an |
| 1034 | /// appropriate form and append it to the given result list. |
| 1035 | /// |
| 1036 | /// If the type `T` is based on a higher order value, consider using |
| 1037 | /// `ProcessPDLValueBasedOn` as a base class of the specialization to simplify |
| 1038 | /// the implementation. |
| 1039 | /// |
| 1040 | template <typename T, typename Enable = void> |
| 1041 | struct ProcessPDLValue; |
| 1042 | |
| 1043 | /// This struct provides a simplified model for processing types that are based |
| 1044 | /// on another type, e.g. APInt is based on the handling for IntegerAttr. This |
| 1045 | /// allows for building the necessary processing functions on top of the base |
| 1046 | /// value instead of a PDLValue. Derived users should implement the following |
| 1047 | /// (which subsume the ProcessPDLValue variants): |
| 1048 | /// |
| 1049 | /// static LogicalResult verifyAsArg( |
| 1050 | /// function_ref<LogicalResult(const Twine &)> errorFn, |
| 1051 | /// const BaseT &baseValue, size_t argIdx); |
| 1052 | /// |
| 1053 | /// * This method verifies that the given PDLValue is valid for use as a |
| 1054 | /// value of `T`. |
| 1055 | /// |
| 1056 | /// static T processAsArg(BaseT baseValue); |
| 1057 | /// |
| 1058 | /// * This method processes the given base value as a value of `T`. |
| 1059 | /// |
| 1060 | template <typename T, typename BaseT> |
| 1061 | struct ProcessPDLValueBasedOn { |
| 1062 | static LogicalResult |
| 1063 | verifyAsArg(function_ref<LogicalResult(const Twine &)> errorFn, |
| 1064 | PDLValue pdlValue, size_t argIdx) { |
| 1065 | // Verify the base class before continuing. |
| 1066 | if (failed(ProcessPDLValue<BaseT>::verifyAsArg(errorFn, pdlValue, argIdx))) |
| 1067 | return failure(); |
| 1068 | return ProcessPDLValue<T>::verifyAsArg( |
| 1069 | errorFn, ProcessPDLValue<BaseT>::processAsArg(pdlValue), argIdx); |
| 1070 | } |
| 1071 | static T processAsArg(PDLValue pdlValue) { |
| 1072 | return ProcessPDLValue<T>::processAsArg( |
| 1073 | ProcessPDLValue<BaseT>::processAsArg(pdlValue)); |
| 1074 | } |
| 1075 | |
| 1076 | /// Explicitly add the expected parent API to ensure the parent class |
| 1077 | /// implements the necessary API (and doesn't implicitly inherit it from |
| 1078 | /// somewhere else). |
| 1079 | static LogicalResult |
| 1080 | verifyAsArg(function_ref<LogicalResult(const Twine &)> errorFn, BaseT value, |
| 1081 | size_t argIdx) { |
| 1082 | return success(); |
| 1083 | } |
| 1084 | static T processAsArg(BaseT baseValue); |
| 1085 | }; |
| 1086 | |
| 1087 | /// This struct provides a simplified model for processing types that have |
| 1088 | /// "builtin" PDLValue support: |
| 1089 | /// * Attribute, Operation *, Type, TypeRange, ValueRange |
| 1090 | template <typename T> |
| 1091 | struct ProcessBuiltinPDLValue { |
| 1092 | static LogicalResult |
| 1093 | verifyAsArg(function_ref<LogicalResult(const Twine &)> errorFn, |
| 1094 | PDLValue pdlValue, size_t argIdx) { |
| 1095 | if (pdlValue) |
| 1096 | return success(); |
| 1097 | return errorFn("expected a non-null value for argument " + Twine(argIdx) + |
| 1098 | " of type: " + llvm::getTypeName<T>()); |
| 1099 | } |
| 1100 | |
| 1101 | static T processAsArg(PDLValue pdlValue) { return pdlValue.cast<T>(); } |
| 1102 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1103 | T value) { |
| 1104 | results.push_back(value); |
| 1105 | } |
| 1106 | }; |
| 1107 | |
| 1108 | /// This struct provides a simplified model for processing types that inherit |
| 1109 | /// from builtin PDLValue types. For example, derived attributes like |
| 1110 | /// IntegerAttr, derived types like IntegerType, derived operations like |
| 1111 | /// ModuleOp, Interfaces, etc. |
| 1112 | template <typename T, typename BaseT> |
| 1113 | struct ProcessDerivedPDLValue : public ProcessPDLValueBasedOn<T, BaseT> { |
| 1114 | static LogicalResult |
| 1115 | verifyAsArg(function_ref<LogicalResult(const Twine &)> errorFn, |
| 1116 | BaseT baseValue, size_t argIdx) { |
| 1117 | return TypeSwitch<BaseT, LogicalResult>(baseValue) |
| 1118 | .Case([&](T) { return success(); }) |
| 1119 | .Default([&](BaseT) { |
| 1120 | return errorFn("expected argument " + Twine(argIdx) + |
| 1121 | " to be of type: " + llvm::getTypeName<T>()); |
| 1122 | }); |
| 1123 | } |
| 1124 | using ProcessPDLValueBasedOn<T, BaseT>::verifyAsArg; |
| 1125 | |
| 1126 | static T processAsArg(BaseT baseValue) { |
| 1127 | return baseValue.template cast<T>(); |
| 1128 | } |
| 1129 | using ProcessPDLValueBasedOn<T, BaseT>::processAsArg; |
| 1130 | |
| 1131 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1132 | T value) { |
| 1133 | results.push_back(value); |
| 1134 | } |
| 1135 | }; |
| 1136 | |
| 1137 | //===----------------------------------------------------------------------===// |
| 1138 | // Attribute |
| 1139 | |
| 1140 | template <> |
| 1141 | struct ProcessPDLValue<Attribute> : public ProcessBuiltinPDLValue<Attribute> {}; |
| 1142 | template <typename T> |
| 1143 | struct ProcessPDLValue<T, |
| 1144 | std::enable_if_t<std::is_base_of<Attribute, T>::value>> |
| 1145 | : public ProcessDerivedPDLValue<T, Attribute> {}; |
| 1146 | |
| 1147 | /// Handling for various Attribute value types. |
| 1148 | template <> |
| 1149 | struct ProcessPDLValue<StringRef> |
| 1150 | : public ProcessPDLValueBasedOn<StringRef, StringAttr> { |
| 1151 | static StringRef processAsArg(StringAttr value) { return value.getValue(); } |
| 1152 | using ProcessPDLValueBasedOn<StringRef, StringAttr>::processAsArg; |
| 1153 | |
| 1154 | static void processAsResult(PatternRewriter &rewriter, PDLResultList &results, |
| 1155 | StringRef value) { |
| 1156 | results.push_back(rewriter.getStringAttr(value)); |
| 1157 | } |
| 1158 | }; |
| 1159 | template <> |
| 1160 | struct ProcessPDLValue<std::string> |
| 1161 | : public ProcessPDLValueBasedOn<std::string, StringAttr> { |
| 1162 | template <typename T> |
| 1163 | static std::string processAsArg(T value) { |
| 1164 | static_assert(always_false<T>, |
| 1165 | "`std::string` arguments require a string copy, use " |
| 1166 | "`StringRef` for string-like arguments instead"); |
| 1167 | return {}; |
| 1168 | } |
| 1169 | static void processAsResult(PatternRewriter &rewriter, PDLResultList &results, |
| 1170 | StringRef value) { |
| 1171 | results.push_back(rewriter.getStringAttr(value)); |
| 1172 | } |
| 1173 | }; |
| 1174 | |
| 1175 | //===----------------------------------------------------------------------===// |
| 1176 | // Operation |
| 1177 | |
| 1178 | template <> |
| 1179 | struct ProcessPDLValue<Operation *> |
| 1180 | : public ProcessBuiltinPDLValue<Operation *> {}; |
| 1181 | template <typename T> |
| 1182 | struct ProcessPDLValue<T, std::enable_if_t<std::is_base_of<OpState, T>::value>> |
| 1183 | : public ProcessDerivedPDLValue<T, Operation *> { |
| 1184 | static T processAsArg(Operation *value) { return cast<T>(value); } |
| 1185 | }; |
| 1186 | |
| 1187 | //===----------------------------------------------------------------------===// |
| 1188 | // Type |
| 1189 | |
| 1190 | template <> |
| 1191 | struct ProcessPDLValue<Type> : public ProcessBuiltinPDLValue<Type> {}; |
| 1192 | template <typename T> |
| 1193 | struct ProcessPDLValue<T, std::enable_if_t<std::is_base_of<Type, T>::value>> |
| 1194 | : public ProcessDerivedPDLValue<T, Type> {}; |
| 1195 | |
| 1196 | //===----------------------------------------------------------------------===// |
| 1197 | // TypeRange |
| 1198 | |
| 1199 | template <> |
| 1200 | struct ProcessPDLValue<TypeRange> : public ProcessBuiltinPDLValue<TypeRange> {}; |
| 1201 | template <> |
| 1202 | struct ProcessPDLValue<ValueTypeRange<OperandRange>> { |
| 1203 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1204 | ValueTypeRange<OperandRange> types) { |
| 1205 | results.push_back(types); |
| 1206 | } |
| 1207 | }; |
| 1208 | template <> |
| 1209 | struct ProcessPDLValue<ValueTypeRange<ResultRange>> { |
| 1210 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1211 | ValueTypeRange<ResultRange> types) { |
| 1212 | results.push_back(types); |
| 1213 | } |
| 1214 | }; |
| 1215 | template <unsigned N> |
| 1216 | struct ProcessPDLValue<SmallVector<Type, N>> { |
| 1217 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1218 | SmallVector<Type, N> values) { |
| 1219 | results.push_back(TypeRange(values)); |
| 1220 | } |
| 1221 | }; |
| 1222 | |
| 1223 | //===----------------------------------------------------------------------===// |
| 1224 | // Value |
| 1225 | |
| 1226 | template <> |
| 1227 | struct ProcessPDLValue<Value> : public ProcessBuiltinPDLValue<Value> {}; |
| 1228 | |
| 1229 | //===----------------------------------------------------------------------===// |
| 1230 | // ValueRange |
| 1231 | |
| 1232 | template <> |
| 1233 | struct ProcessPDLValue<ValueRange> : public ProcessBuiltinPDLValue<ValueRange> { |
| 1234 | }; |
| 1235 | template <> |
| 1236 | struct ProcessPDLValue<OperandRange> { |
| 1237 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1238 | OperandRange values) { |
| 1239 | results.push_back(values); |
| 1240 | } |
| 1241 | }; |
| 1242 | template <> |
| 1243 | struct ProcessPDLValue<ResultRange> { |
| 1244 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1245 | ResultRange values) { |
| 1246 | results.push_back(values); |
| 1247 | } |
| 1248 | }; |
| 1249 | template <unsigned N> |
| 1250 | struct ProcessPDLValue<SmallVector<Value, N>> { |
| 1251 | static void processAsResult(PatternRewriter &, PDLResultList &results, |
| 1252 | SmallVector<Value, N> values) { |
| 1253 | results.push_back(ValueRange(values)); |
| 1254 | } |
| 1255 | }; |
| 1256 | |
| 1257 | //===----------------------------------------------------------------------===// |
| 1258 | // PDL Function Builder: Argument Handling |
| 1259 | //===----------------------------------------------------------------------===// |
| 1260 | |
| 1261 | /// Validate the given PDLValues match the constraints defined by the argument |
| 1262 | /// types of the given function. In the case of failure, a match failure |
| 1263 | /// diagnostic is emitted. |
| 1264 | /// FIXME: This should be completely removed in favor of `assertArgs`, but PDL |
| 1265 | /// does not currently preserve Constraint application ordering. |
| 1266 | template <typename PDLFnT, std::size_t... I> |
| 1267 | LogicalResult verifyAsArgs(PatternRewriter &rewriter, ArrayRef<PDLValue> values, |
| 1268 | std::index_sequence<I...>) { |
| 1269 | using FnTraitsT = llvm::function_traits<PDLFnT>; |
| 1270 | |
| 1271 | auto errorFn = [&](const Twine &msg) { |
| 1272 | return rewriter.notifyMatchFailure(rewriter.getUnknownLoc(), msg); |
| 1273 | }; |
| 1274 | return success( |
| 1275 | (succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: |
| 1276 | verifyAsArg(errorFn, values[I], I)) && |
| 1277 | ...)); |
| 1278 | } |
| 1279 | |
| 1280 | /// Assert that the given PDLValues match the constraints defined by the |
| 1281 | /// arguments of the given function. In the case of failure, a fatal error |
| 1282 | /// is emitted. |
| 1283 | template <typename PDLFnT, std::size_t... I> |
| 1284 | void assertArgs(PatternRewriter &rewriter, ArrayRef<PDLValue> values, |
| 1285 | std::index_sequence<I...>) { |
| 1286 | // We only want to do verification in debug builds, same as with `assert`. |
| 1287 | #if LLVM_ENABLE_ABI_BREAKING_CHECKS1 |
| 1288 | using FnTraitsT = llvm::function_traits<PDLFnT>; |
| 1289 | auto errorFn = [&](const Twine &msg) -> LogicalResult { |
| 1290 | llvm::report_fatal_error(msg); |
| 1291 | }; |
| 1292 | (void)errorFn; |
| 1293 | assert((succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>::(static_cast <bool> ((succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn , values[I], I)) && ...)) ? void (0) : __assert_fail ( "(succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn, values[I], I)) && ...)" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 1295, __extension__ __PRETTY_FUNCTION__)) |
| 1294 | verifyAsArg(errorFn, values[I], I)) &&(static_cast <bool> ((succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn , values[I], I)) && ...)) ? void (0) : __assert_fail ( "(succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn, values[I], I)) && ...)" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 1295, __extension__ __PRETTY_FUNCTION__)) |
| 1295 | ...))(static_cast <bool> ((succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn , values[I], I)) && ...)) ? void (0) : __assert_fail ( "(succeeded(ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: verifyAsArg(errorFn, values[I], I)) && ...)" , "llvm/../mlir/include/mlir/IR/PatternMatch.h", 1295, __extension__ __PRETTY_FUNCTION__)); |
| 1296 | #endif |
| 1297 | (void)values; |
| 1298 | } |
| 1299 | |
| 1300 | //===----------------------------------------------------------------------===// |
| 1301 | // PDL Function Builder: Results Handling |
| 1302 | //===----------------------------------------------------------------------===// |
| 1303 | |
| 1304 | /// Store a single result within the result list. |
| 1305 | template <typename T> |
| 1306 | static LogicalResult processResults(PatternRewriter &rewriter, |
| 1307 | PDLResultList &results, T &&value) { |
| 1308 | ProcessPDLValue<T>::processAsResult(rewriter, results, |
| 1309 | std::forward<T>(value)); |
| 1310 | return success(); |
| 1311 | } |
| 1312 | |
| 1313 | /// Store a std::pair<> as individual results within the result list. |
| 1314 | template <typename T1, typename T2> |
| 1315 | static LogicalResult processResults(PatternRewriter &rewriter, |
| 1316 | PDLResultList &results, |
| 1317 | std::pair<T1, T2> &&pair) { |
| 1318 | if (failed(processResults(rewriter, results, std::move(pair.first))) || |
| 1319 | failed(processResults(rewriter, results, std::move(pair.second)))) |
| 1320 | return failure(); |
| 1321 | return success(); |
| 1322 | } |
| 1323 | |
| 1324 | /// Store a std::tuple<> as individual results within the result list. |
| 1325 | template <typename... Ts> |
| 1326 | static LogicalResult processResults(PatternRewriter &rewriter, |
| 1327 | PDLResultList &results, |
| 1328 | std::tuple<Ts...> &&tuple) { |
| 1329 | auto applyFn = [&](auto &&...args) { |
| 1330 | return (succeeded(processResults(rewriter, results, std::move(args))) && |
| 1331 | ...); |
| 1332 | }; |
| 1333 | return success(std::apply(applyFn, std::move(tuple))); |
| 1334 | } |
| 1335 | |
| 1336 | /// Handle LogicalResult propagation. |
| 1337 | inline LogicalResult processResults(PatternRewriter &rewriter, |
| 1338 | PDLResultList &results, |
| 1339 | LogicalResult &&result) { |
| 1340 | return result; |
| 1341 | } |
| 1342 | template <typename T> |
| 1343 | static LogicalResult processResults(PatternRewriter &rewriter, |
| 1344 | PDLResultList &results, |
| 1345 | FailureOr<T> &&result) { |
| 1346 | if (failed(result)) |
| 1347 | return failure(); |
| 1348 | return processResults(rewriter, results, std::move(*result)); |
| 1349 | } |
| 1350 | |
| 1351 | //===----------------------------------------------------------------------===// |
| 1352 | // PDL Constraint Builder |
| 1353 | //===----------------------------------------------------------------------===// |
| 1354 | |
| 1355 | /// Process the arguments of a native constraint and invoke it. |
| 1356 | template <typename PDLFnT, std::size_t... I, |
| 1357 | typename FnTraitsT = llvm::function_traits<PDLFnT>> |
| 1358 | typename FnTraitsT::result_t |
| 1359 | processArgsAndInvokeConstraint(PDLFnT &fn, PatternRewriter &rewriter, |
| 1360 | ArrayRef<PDLValue> values, |
| 1361 | std::index_sequence<I...>) { |
| 1362 | return fn( |
| 1363 | rewriter, |
| 1364 | (ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>::processAsArg( |
| 1365 | values[I]))...); |
| 1366 | } |
| 1367 | |
| 1368 | /// Build a constraint function from the given function `ConstraintFnT`. This |
| 1369 | /// allows for enabling the user to define simpler, more direct constraint |
| 1370 | /// functions without needing to handle the low-level PDL goop. |
| 1371 | /// |
| 1372 | /// If the constraint function is already in the correct form, we just forward |
| 1373 | /// it directly. |
| 1374 | template <typename ConstraintFnT> |
| 1375 | std::enable_if_t< |
| 1376 | std::is_convertible<ConstraintFnT, PDLConstraintFunction>::value, |
| 1377 | PDLConstraintFunction> |
| 1378 | buildConstraintFn(ConstraintFnT &&constraintFn) { |
| 1379 | return std::forward<ConstraintFnT>(constraintFn); |
| 1380 | } |
| 1381 | /// Otherwise, we generate a wrapper that will unpack the PDLValues in the form |
| 1382 | /// we desire. |
| 1383 | template <typename ConstraintFnT> |
| 1384 | std::enable_if_t< |
| 1385 | !std::is_convertible<ConstraintFnT, PDLConstraintFunction>::value, |
| 1386 | PDLConstraintFunction> |
| 1387 | buildConstraintFn(ConstraintFnT &&constraintFn) { |
| 1388 | return [constraintFn = std::forward<ConstraintFnT>(constraintFn)]( |
| 1389 | PatternRewriter &rewriter, |
| 1390 | ArrayRef<PDLValue> values) -> LogicalResult { |
| 1391 | auto argIndices = std::make_index_sequence< |
| 1392 | llvm::function_traits<ConstraintFnT>::num_args - 1>(); |
| 1393 | if (failed(verifyAsArgs<ConstraintFnT>(rewriter, values, argIndices))) |
| 1394 | return failure(); |
| 1395 | return processArgsAndInvokeConstraint(constraintFn, rewriter, values, |
| 1396 | argIndices); |
| 1397 | }; |
| 1398 | } |
| 1399 | |
| 1400 | //===----------------------------------------------------------------------===// |
| 1401 | // PDL Rewrite Builder |
| 1402 | //===----------------------------------------------------------------------===// |
| 1403 | |
| 1404 | /// Process the arguments of a native rewrite and invoke it. |
| 1405 | /// This overload handles the case of no return values. |
| 1406 | template <typename PDLFnT, std::size_t... I, |
| 1407 | typename FnTraitsT = llvm::function_traits<PDLFnT>> |
| 1408 | std::enable_if_t<std::is_same<typename FnTraitsT::result_t, void>::value, |
| 1409 | LogicalResult> |
| 1410 | processArgsAndInvokeRewrite(PDLFnT &fn, PatternRewriter &rewriter, |
| 1411 | PDLResultList &, ArrayRef<PDLValue> values, |
| 1412 | std::index_sequence<I...>) { |
| 1413 | fn(rewriter, |
| 1414 | (ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>::processAsArg( |
| 1415 | values[I]))...); |
| 1416 | return success(); |
| 1417 | } |
| 1418 | /// This overload handles the case of return values, which need to be packaged |
| 1419 | /// into the result list. |
| 1420 | template <typename PDLFnT, std::size_t... I, |
| 1421 | typename FnTraitsT = llvm::function_traits<PDLFnT>> |
| 1422 | std::enable_if_t<!std::is_same<typename FnTraitsT::result_t, void>::value, |
| 1423 | LogicalResult> |
| 1424 | processArgsAndInvokeRewrite(PDLFnT &fn, PatternRewriter &rewriter, |
| 1425 | PDLResultList &results, ArrayRef<PDLValue> values, |
| 1426 | std::index_sequence<I...>) { |
| 1427 | return processResults( |
| 1428 | rewriter, results, |
| 1429 | fn(rewriter, (ProcessPDLValue<typename FnTraitsT::template arg_t<I + 1>>:: |
| 1430 | processAsArg(values[I]))...)); |
| 1431 | (void)values; |
| 1432 | } |
| 1433 | |
| 1434 | /// Build a rewrite function from the given function `RewriteFnT`. This |
| 1435 | /// allows for enabling the user to define simpler, more direct rewrite |
| 1436 | /// functions without needing to handle the low-level PDL goop. |
| 1437 | /// |
| 1438 | /// If the rewrite function is already in the correct form, we just forward |
| 1439 | /// it directly. |
| 1440 | template <typename RewriteFnT> |
| 1441 | std::enable_if_t<std::is_convertible<RewriteFnT, PDLRewriteFunction>::value, |
| 1442 | PDLRewriteFunction> |
| 1443 | buildRewriteFn(RewriteFnT &&rewriteFn) { |
| 1444 | return std::forward<RewriteFnT>(rewriteFn); |
| 1445 | } |
| 1446 | /// Otherwise, we generate a wrapper that will unpack the PDLValues in the form |
| 1447 | /// we desire. |
| 1448 | template <typename RewriteFnT> |
| 1449 | std::enable_if_t<!std::is_convertible<RewriteFnT, PDLRewriteFunction>::value, |
| 1450 | PDLRewriteFunction> |
| 1451 | buildRewriteFn(RewriteFnT &&rewriteFn) { |
| 1452 | return [rewriteFn = std::forward<RewriteFnT>(rewriteFn)]( |
| 1453 | PatternRewriter &rewriter, PDLResultList &results, |
| 1454 | ArrayRef<PDLValue> values) { |
| 1455 | auto argIndices = |
| 1456 | std::make_index_sequence<llvm::function_traits<RewriteFnT>::num_args - |
| 1457 | 1>(); |
| 1458 | assertArgs<RewriteFnT>(rewriter, values, argIndices); |
| 1459 | return processArgsAndInvokeRewrite(rewriteFn, rewriter, results, values, |
| 1460 | argIndices); |
| 1461 | }; |
| 1462 | } |
| 1463 | |
| 1464 | } // namespace pdl_function_builder |
| 1465 | } // namespace detail |
| 1466 | |
| 1467 | //===----------------------------------------------------------------------===// |
| 1468 | // PDLPatternModule |
| 1469 | |
| 1470 | /// This class contains all of the necessary data for a set of PDL patterns, or |
| 1471 | /// pattern rewrites specified in the form of the PDL dialect. This PDL module |
| 1472 | /// contained by this pattern may contain any number of `pdl.pattern` |
| 1473 | /// operations. |
| 1474 | class PDLPatternModule { |
| 1475 | public: |
| 1476 | PDLPatternModule() = default; |
| 1477 | |
| 1478 | /// Construct a PDL pattern with the given module and configurations. |
| 1479 | PDLPatternModule(OwningOpRef<ModuleOp> module) |
| 1480 | : pdlModule(std::move(module)) {} |
| 1481 | template <typename... ConfigsT> |
| 1482 | PDLPatternModule(OwningOpRef<ModuleOp> module, ConfigsT &&...patternConfigs) |
| 1483 | : PDLPatternModule(std::move(module)) { |
| 1484 | auto configSet = std::make_unique<PDLPatternConfigSet>( |
| 1485 | std::forward<ConfigsT>(patternConfigs)...); |
| 1486 | attachConfigToPatterns(*pdlModule, *configSet); |
| 1487 | configs.emplace_back(std::move(configSet)); |
| 1488 | } |
| 1489 | |
| 1490 | /// Merge the state in `other` into this pattern module. |
| 1491 | void mergeIn(PDLPatternModule &&other); |
| 1492 | |
| 1493 | /// Return the internal PDL module of this pattern. |
| 1494 | ModuleOp getModule() { return pdlModule.get(); } |
| 1495 | |
| 1496 | //===--------------------------------------------------------------------===// |
| 1497 | // Function Registry |
| 1498 | |
| 1499 | /// Register a constraint function with PDL. A constraint function may be |
| 1500 | /// specified in one of two ways: |
| 1501 | /// |
| 1502 | /// * `LogicalResult (PatternRewriter &, ArrayRef<PDLValue>)` |
| 1503 | /// |
| 1504 | /// In this overload the arguments of the constraint function are passed via |
| 1505 | /// the low-level PDLValue form. |
| 1506 | /// |
| 1507 | /// * `LogicalResult (PatternRewriter &, ValueTs... values)` |
| 1508 | /// |
| 1509 | /// In this form the arguments of the constraint function are passed via the |
| 1510 | /// expected high level C++ type. In this form, the framework will |
| 1511 | /// automatically unwrap PDLValues and convert them to the expected ValueTs. |
| 1512 | /// For example, if the constraint function accepts a `Operation *`, the |
| 1513 | /// framework will automatically cast the input PDLValue. In the case of a |
| 1514 | /// `StringRef`, the framework will automatically unwrap the argument as a |
| 1515 | /// StringAttr and pass the underlying string value. To see the full list of |
| 1516 | /// supported types, or to see how to add handling for custom types, view |
| 1517 | /// the definition of `ProcessPDLValue` above. |
| 1518 | void registerConstraintFunction(StringRef name, |
| 1519 | PDLConstraintFunction constraintFn); |
| 1520 | template <typename ConstraintFnT> |
| 1521 | void registerConstraintFunction(StringRef name, |
| 1522 | ConstraintFnT &&constraintFn) { |
| 1523 | registerConstraintFunction(name, |
| 1524 | detail::pdl_function_builder::buildConstraintFn( |
| 1525 | std::forward<ConstraintFnT>(constraintFn))); |
| 1526 | } |
| 1527 | |
| 1528 | /// Register a rewrite function with PDL. A rewrite function may be specified |
| 1529 | /// in one of two ways: |
| 1530 | /// |
| 1531 | /// * `void (PatternRewriter &, PDLResultList &, ArrayRef<PDLValue>)` |
| 1532 | /// |
| 1533 | /// In this overload the arguments of the constraint function are passed via |
| 1534 | /// the low-level PDLValue form, and the results are manually appended to |
| 1535 | /// the given result list. |
| 1536 | /// |
| 1537 | /// * `ResultT (PatternRewriter &, ValueTs... values)` |
| 1538 | /// |
| 1539 | /// In this form the arguments and result of the rewrite function are passed |
| 1540 | /// via the expected high level C++ type. In this form, the framework will |
| 1541 | /// automatically unwrap the PDLValues arguments and convert them to the |
| 1542 | /// expected ValueTs. It will also automatically handle the processing and |
| 1543 | /// packaging of the result value to the result list. For example, if the |
| 1544 | /// rewrite function takes a `Operation *`, the framework will automatically |
| 1545 | /// cast the input PDLValue. In the case of a `StringRef`, the framework |
| 1546 | /// will automatically unwrap the argument as a StringAttr and pass the |
| 1547 | /// underlying string value. In the reverse case, if the rewrite returns a |
| 1548 | /// StringRef or std::string, it will automatically package this as a |
| 1549 | /// StringAttr and append it to the result list. To see the full list of |
| 1550 | /// supported types, or to see how to add handling for custom types, view |
| 1551 | /// the definition of `ProcessPDLValue` above. |
| 1552 | void registerRewriteFunction(StringRef name, PDLRewriteFunction rewriteFn); |
| 1553 | template <typename RewriteFnT> |
| 1554 | void registerRewriteFunction(StringRef name, RewriteFnT &&rewriteFn) { |
| 1555 | registerRewriteFunction(name, detail::pdl_function_builder::buildRewriteFn( |
| 1556 | std::forward<RewriteFnT>(rewriteFn))); |
| 1557 | } |
| 1558 | |
| 1559 | /// Return the set of the registered constraint functions. |
| 1560 | const llvm::StringMap<PDLConstraintFunction> &getConstraintFunctions() const { |
| 1561 | return constraintFunctions; |
| 1562 | } |
| 1563 | llvm::StringMap<PDLConstraintFunction> takeConstraintFunctions() { |
| 1564 | return constraintFunctions; |
| 1565 | } |
| 1566 | /// Return the set of the registered rewrite functions. |
| 1567 | const llvm::StringMap<PDLRewriteFunction> &getRewriteFunctions() const { |
| 1568 | return rewriteFunctions; |
| 1569 | } |
| 1570 | llvm::StringMap<PDLRewriteFunction> takeRewriteFunctions() { |
| 1571 | return rewriteFunctions; |
| 1572 | } |
| 1573 | |
| 1574 | /// Return the set of the registered pattern configs. |
| 1575 | SmallVector<std::unique_ptr<PDLPatternConfigSet>> takeConfigs() { |
| 1576 | return std::move(configs); |
| 1577 | } |
| 1578 | DenseMap<Operation *, PDLPatternConfigSet *> takeConfigMap() { |
| 1579 | return std::move(configMap); |
| 1580 | } |
| 1581 | |
| 1582 | /// Clear out the patterns and functions within this module. |
| 1583 | void clear() { |
| 1584 | pdlModule = nullptr; |
| 1585 | constraintFunctions.clear(); |
| 1586 | rewriteFunctions.clear(); |
| 1587 | } |
| 1588 | |
| 1589 | private: |
| 1590 | /// Attach the given pattern config set to the patterns defined within the |
| 1591 | /// given module. |
| 1592 | void attachConfigToPatterns(ModuleOp module, PDLPatternConfigSet &configSet); |
| 1593 | |
| 1594 | /// The module containing the `pdl.pattern` operations. |
| 1595 | OwningOpRef<ModuleOp> pdlModule; |
| 1596 | |
| 1597 | /// The set of configuration sets referenced by patterns within `pdlModule`. |
| 1598 | SmallVector<std::unique_ptr<PDLPatternConfigSet>> configs; |
| 1599 | DenseMap<Operation *, PDLPatternConfigSet *> configMap; |
| 1600 | |
| 1601 | /// The external functions referenced from within the PDL module. |
| 1602 | llvm::StringMap<PDLConstraintFunction> constraintFunctions; |
| 1603 | llvm::StringMap<PDLRewriteFunction> rewriteFunctions; |
| 1604 | }; |
| 1605 | |
| 1606 | //===----------------------------------------------------------------------===// |
| 1607 | // RewritePatternSet |
| 1608 | //===----------------------------------------------------------------------===// |
| 1609 | |
| 1610 | class RewritePatternSet { |
| 1611 | using NativePatternListT = std::vector<std::unique_ptr<RewritePattern>>; |
| 1612 | |
| 1613 | public: |
| 1614 | RewritePatternSet(MLIRContext *context) : context(context) {} |
| 1615 | |
| 1616 | /// Construct a RewritePatternSet populated with the given pattern. |
| 1617 | RewritePatternSet(MLIRContext *context, |
| 1618 | std::unique_ptr<RewritePattern> pattern) |
| 1619 | : context(context) { |
| 1620 | nativePatterns.emplace_back(std::move(pattern)); |
| 1621 | } |
| 1622 | RewritePatternSet(PDLPatternModule &&pattern) |
| 1623 | : context(pattern.getModule()->getContext()), |
| 1624 | pdlPatterns(std::move(pattern)) {} |
| 1625 | |
| 1626 | MLIRContext *getContext() const { return context; } |
| 1627 | |
| 1628 | /// Return the native patterns held in this list. |
| 1629 | NativePatternListT &getNativePatterns() { return nativePatterns; } |
| 1630 | |
| 1631 | /// Return the PDL patterns held in this list. |
| 1632 | PDLPatternModule &getPDLPatterns() { return pdlPatterns; } |
| 1633 | |
| 1634 | /// Clear out all of the held patterns in this list. |
| 1635 | void clear() { |
| 1636 | nativePatterns.clear(); |
| 1637 | pdlPatterns.clear(); |
| 1638 | } |
| 1639 | |
| 1640 | //===--------------------------------------------------------------------===// |
| 1641 | // 'add' methods for adding patterns to the set. |
| 1642 | //===--------------------------------------------------------------------===// |
| 1643 | |
| 1644 | /// Add an instance of each of the pattern types 'Ts' to the pattern list with |
| 1645 | /// the given arguments. Return a reference to `this` for chaining insertions. |
| 1646 | /// Note: ConstructorArg is necessary here to separate the two variadic lists. |
| 1647 | template <typename... Ts, typename ConstructorArg, |
| 1648 | typename... ConstructorArgs, |
| 1649 | typename = std::enable_if_t<sizeof...(Ts) != 0>> |
| 1650 | RewritePatternSet &add(ConstructorArg &&arg, ConstructorArgs &&...args) { |
| 1651 | // The following expands a call to emplace_back for each of the pattern |
| 1652 | // types 'Ts'. |
| 1653 | (addImpl<Ts>(/*debugLabels=*/std::nullopt, |
| 1654 | std::forward<ConstructorArg>(arg), |
| 1655 | std::forward<ConstructorArgs>(args)...), |
| 1656 | ...); |
| 1657 | return *this; |
| 1658 | } |
| 1659 | /// An overload of the above `add` method that allows for attaching a set |
| 1660 | /// of debug labels to the attached patterns. This is useful for labeling |
| 1661 | /// groups of patterns that may be shared between multiple different |
| 1662 | /// passes/users. |
| 1663 | template <typename... Ts, typename ConstructorArg, |
| 1664 | typename... ConstructorArgs, |
| 1665 | typename = std::enable_if_t<sizeof...(Ts) != 0>> |
| 1666 | RewritePatternSet &addWithLabel(ArrayRef<StringRef> debugLabels, |
| 1667 | ConstructorArg &&arg, |
| 1668 | ConstructorArgs &&...args) { |
| 1669 | // The following expands a call to emplace_back for each of the pattern |
| 1670 | // types 'Ts'. |
| 1671 | (addImpl<Ts>(debugLabels, arg, args...), ...); |
| 1672 | return *this; |
| 1673 | } |
| 1674 | |
| 1675 | /// Add an instance of each of the pattern types 'Ts'. Return a reference to |
| 1676 | /// `this` for chaining insertions. |
| 1677 | template <typename... Ts> |
| 1678 | RewritePatternSet &add() { |
| 1679 | (addImpl<Ts>(), ...); |
| 1680 | return *this; |
| 1681 | } |
| 1682 | |
| 1683 | /// Add the given native pattern to the pattern list. Return a reference to |
| 1684 | /// `this` for chaining insertions. |
| 1685 | RewritePatternSet &add(std::unique_ptr<RewritePattern> pattern) { |
| 1686 | nativePatterns.emplace_back(std::move(pattern)); |
| 1687 | return *this; |
| 1688 | } |
| 1689 | |
| 1690 | /// Add the given PDL pattern to the pattern list. Return a reference to |
| 1691 | /// `this` for chaining insertions. |
| 1692 | RewritePatternSet &add(PDLPatternModule &&pattern) { |
| 1693 | pdlPatterns.mergeIn(std::move(pattern)); |
| 1694 | return *this; |
| 1695 | } |
| 1696 | |
| 1697 | // Add a matchAndRewrite style pattern represented as a C function pointer. |
| 1698 | template <typename OpType> |
| 1699 | RewritePatternSet & |
| 1700 | add(LogicalResult (*implFn)(OpType, PatternRewriter &rewriter), |
| 1701 | PatternBenefit benefit = 1, ArrayRef<StringRef> generatedNames = {}) { |
| 1702 | struct FnPattern final : public OpRewritePattern<OpType> { |
| 1703 | FnPattern(LogicalResult (*implFn)(OpType, PatternRewriter &rewriter), |
| 1704 | MLIRContext *context, PatternBenefit benefit, |
| 1705 | ArrayRef<StringRef> generatedNames) |
| 1706 | : OpRewritePattern<OpType>(context, benefit, generatedNames), |
| 1707 | implFn(implFn) {} |
| 1708 | |
| 1709 | LogicalResult matchAndRewrite(OpType op, |
| 1710 | PatternRewriter &rewriter) const override { |
| 1711 | return implFn(op, rewriter); |
| 1712 | } |
| 1713 | |
| 1714 | private: |
| 1715 | LogicalResult (*implFn)(OpType, PatternRewriter &rewriter); |
| 1716 | }; |
| 1717 | add(std::make_unique<FnPattern>(std::move(implFn), getContext(), benefit, |
| 1718 | generatedNames)); |
| 1719 | return *this; |
| 1720 | } |
| 1721 | |
| 1722 | //===--------------------------------------------------------------------===// |
| 1723 | // Pattern Insertion |
| 1724 | //===--------------------------------------------------------------------===// |
| 1725 | |
| 1726 | // TODO: These are soft deprecated in favor of the 'add' methods above. |
| 1727 | |
| 1728 | /// Add an instance of each of the pattern types 'Ts' to the pattern list with |
| 1729 | /// the given arguments. Return a reference to `this` for chaining insertions. |
| 1730 | /// Note: ConstructorArg is necessary here to separate the two variadic lists. |
| 1731 | template <typename... Ts, typename ConstructorArg, |
| 1732 | typename... ConstructorArgs, |
| 1733 | typename = std::enable_if_t<sizeof...(Ts) != 0>> |
| 1734 | RewritePatternSet &insert(ConstructorArg &&arg, ConstructorArgs &&...args) { |
| 1735 | // The following expands a call to emplace_back for each of the pattern |
| 1736 | // types 'Ts'. |
| 1737 | (addImpl<Ts>(/*debugLabels=*/std::nullopt, arg, args...), ...); |
| 1738 | return *this; |
| 1739 | } |
| 1740 | |
| 1741 | /// Add an instance of each of the pattern types 'Ts'. Return a reference to |
| 1742 | /// `this` for chaining insertions. |
| 1743 | template <typename... Ts> |
| 1744 | RewritePatternSet &insert() { |
| 1745 | (addImpl<Ts>(), ...); |
| 1746 | return *this; |
| 1747 | } |
| 1748 | |
| 1749 | /// Add the given native pattern to the pattern list. Return a reference to |
| 1750 | /// `this` for chaining insertions. |
| 1751 | RewritePatternSet &insert(std::unique_ptr<RewritePattern> pattern) { |
| 1752 | nativePatterns.emplace_back(std::move(pattern)); |
| 1753 | return *this; |
| 1754 | } |
| 1755 | |
| 1756 | /// Add the given PDL pattern to the pattern list. Return a reference to |
| 1757 | /// `this` for chaining insertions. |
| 1758 | RewritePatternSet &insert(PDLPatternModule &&pattern) { |
| 1759 | pdlPatterns.mergeIn(std::move(pattern)); |
| 1760 | return *this; |
| 1761 | } |
| 1762 | |
| 1763 | // Add a matchAndRewrite style pattern represented as a C function pointer. |
| 1764 | template <typename OpType> |
| 1765 | RewritePatternSet & |
| 1766 | insert(LogicalResult (*implFn)(OpType, PatternRewriter &rewriter)) { |
| 1767 | struct FnPattern final : public OpRewritePattern<OpType> { |
| 1768 | FnPattern(LogicalResult (*implFn)(OpType, PatternRewriter &rewriter), |
| 1769 | MLIRContext *context) |
| 1770 | : OpRewritePattern<OpType>(context), implFn(implFn) { |
| 1771 | this->setDebugName(llvm::getTypeName<FnPattern>()); |
| 1772 | } |
| 1773 | |
| 1774 | LogicalResult matchAndRewrite(OpType op, |
| 1775 | PatternRewriter &rewriter) const override { |
| 1776 | return implFn(op, rewriter); |
| 1777 | } |
| 1778 | |
| 1779 | private: |
| 1780 | LogicalResult (*implFn)(OpType, PatternRewriter &rewriter); |
| 1781 | }; |
| 1782 | add(std::make_unique<FnPattern>(std::move(implFn), getContext())); |
| 1783 | return *this; |
| 1784 | } |
| 1785 | |
| 1786 | private: |
| 1787 | /// Add an instance of the pattern type 'T'. Return a reference to `this` for |
| 1788 | /// chaining insertions. |
| 1789 | template <typename T, typename... Args> |
| 1790 | std::enable_if_t<std::is_base_of<RewritePattern, T>::value> |
| 1791 | addImpl(ArrayRef<StringRef> debugLabels, Args &&...args) { |
| 1792 | std::unique_ptr<T> pattern = |
| 1793 | RewritePattern::create<T>(std::forward<Args>(args)...); |
| 1794 | pattern->addDebugLabels(debugLabels); |
| 1795 | nativePatterns.emplace_back(std::move(pattern)); |
| 1796 | } |
| 1797 | template <typename T, typename... Args> |
| 1798 | std::enable_if_t<std::is_base_of<PDLPatternModule, T>::value> |
| 1799 | addImpl(ArrayRef<StringRef> debugLabels, Args &&...args) { |
| 1800 | // TODO: Add the provided labels to the PDL pattern when PDL supports |
| 1801 | // labels. |
| 1802 | pdlPatterns.mergeIn(T(std::forward<Args>(args)...)); |
| 1803 | } |
| 1804 | |
| 1805 | MLIRContext *const context; |
| 1806 | NativePatternListT nativePatterns; |
| 1807 | PDLPatternModule pdlPatterns; |
| 1808 | }; |
| 1809 | |
| 1810 | } // namespace mlir |
| 1811 | |
| 1812 | #endif // MLIR_IR_PATTERNMATCH_H |
| 1 | //===- Builders.h - Helpers for constructing MLIR Classes -------*- C++ -*-===// | |||
| 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 | #ifndef MLIR_IR_BUILDERS_H | |||
| 10 | #define MLIR_IR_BUILDERS_H | |||
| 11 | ||||
| 12 | #include "mlir/IR/OpDefinition.h" | |||
| 13 | #include "llvm/Support/Compiler.h" | |||
| 14 | #include <optional> | |||
| 15 | ||||
| 16 | namespace mlir { | |||
| 17 | ||||
| 18 | class AffineExpr; | |||
| 19 | class IRMapping; | |||
| 20 | class UnknownLoc; | |||
| 21 | class FileLineColLoc; | |||
| 22 | class Type; | |||
| 23 | class PrimitiveType; | |||
| 24 | class IntegerType; | |||
| 25 | class FloatType; | |||
| 26 | class FunctionType; | |||
| 27 | class IndexType; | |||
| 28 | class MemRefType; | |||
| 29 | class VectorType; | |||
| 30 | class RankedTensorType; | |||
| 31 | class UnrankedTensorType; | |||
| 32 | class TupleType; | |||
| 33 | class NoneType; | |||
| 34 | class BoolAttr; | |||
| 35 | class IntegerAttr; | |||
| 36 | class FloatAttr; | |||
| 37 | class StringAttr; | |||
| 38 | class TypeAttr; | |||
| 39 | class ArrayAttr; | |||
| 40 | class SymbolRefAttr; | |||
| 41 | class ElementsAttr; | |||
| 42 | class DenseElementsAttr; | |||
| 43 | class DenseIntElementsAttr; | |||
| 44 | class AffineMapAttr; | |||
| 45 | class AffineMap; | |||
| 46 | class UnitAttr; | |||
| 47 | ||||
| 48 | /// This class is a general helper class for creating context-global objects | |||
| 49 | /// like types, attributes, and affine expressions. | |||
| 50 | class Builder { | |||
| 51 | public: | |||
| 52 | explicit Builder(MLIRContext *context) : context(context) {} | |||
| 53 | explicit Builder(Operation *op) : Builder(op->getContext()) {} | |||
| 54 | ||||
| 55 | MLIRContext *getContext() const { return context; } | |||
| 56 | ||||
| 57 | // Locations. | |||
| 58 | Location getUnknownLoc(); | |||
| 59 | Location getFusedLoc(ArrayRef<Location> locs, | |||
| 60 | Attribute metadata = Attribute()); | |||
| 61 | ||||
| 62 | // Types. | |||
| 63 | FloatType getFloat8E5M2Type(); | |||
| 64 | FloatType getFloat8E4M3FNType(); | |||
| 65 | FloatType getFloat8E5M2FNUZType(); | |||
| 66 | FloatType getFloat8E4M3FNUZType(); | |||
| 67 | FloatType getFloat8E4M3B11FNUZType(); | |||
| 68 | FloatType getBF16Type(); | |||
| 69 | FloatType getF16Type(); | |||
| 70 | FloatType getF32Type(); | |||
| 71 | FloatType getF64Type(); | |||
| 72 | FloatType getF80Type(); | |||
| 73 | FloatType getF128Type(); | |||
| 74 | ||||
| 75 | IndexType getIndexType(); | |||
| 76 | ||||
| 77 | IntegerType getI1Type(); | |||
| 78 | IntegerType getI2Type(); | |||
| 79 | IntegerType getI4Type(); | |||
| 80 | IntegerType getI8Type(); | |||
| 81 | IntegerType getI16Type(); | |||
| 82 | IntegerType getI32Type(); | |||
| 83 | IntegerType getI64Type(); | |||
| 84 | IntegerType getIntegerType(unsigned width); | |||
| 85 | IntegerType getIntegerType(unsigned width, bool isSigned); | |||
| 86 | FunctionType getFunctionType(TypeRange inputs, TypeRange results); | |||
| 87 | TupleType getTupleType(TypeRange elementTypes); | |||
| 88 | NoneType getNoneType(); | |||
| 89 | ||||
| 90 | /// Get or construct an instance of the type `Ty` with provided arguments. | |||
| 91 | template <typename Ty, typename... Args> | |||
| 92 | Ty getType(Args &&...args) { | |||
| 93 | return Ty::get(context, std::forward<Args>(args)...); | |||
| 94 | } | |||
| 95 | ||||
| 96 | /// Get or construct an instance of the attribute `Attr` with provided | |||
| 97 | /// arguments. | |||
| 98 | template <typename Attr, typename... Args> | |||
| 99 | Attr getAttr(Args &&...args) { | |||
| 100 | return Attr::get(context, std::forward<Args>(args)...); | |||
| 101 | } | |||
| 102 | ||||
| 103 | // Attributes. | |||
| 104 | NamedAttribute getNamedAttr(StringRef name, Attribute val); | |||
| 105 | ||||
| 106 | UnitAttr getUnitAttr(); | |||
| 107 | BoolAttr getBoolAttr(bool value); | |||
| 108 | DictionaryAttr getDictionaryAttr(ArrayRef<NamedAttribute> value); | |||
| 109 | IntegerAttr getIntegerAttr(Type type, int64_t value); | |||
| 110 | IntegerAttr getIntegerAttr(Type type, const APInt &value); | |||
| 111 | FloatAttr getFloatAttr(Type type, double value); | |||
| 112 | FloatAttr getFloatAttr(Type type, const APFloat &value); | |||
| 113 | StringAttr getStringAttr(const Twine &bytes); | |||
| 114 | ArrayAttr getArrayAttr(ArrayRef<Attribute> value); | |||
| 115 | ||||
| 116 | // Returns a 0-valued attribute of the given `type`. This function only | |||
| 117 | // supports boolean, integer, and 16-/32-/64-bit float types, and vector or | |||
| 118 | // ranked tensor of them. Returns null attribute otherwise. | |||
| 119 | TypedAttr getZeroAttr(Type type); | |||
| 120 | ||||
| 121 | // Convenience methods for fixed types. | |||
| 122 | FloatAttr getF16FloatAttr(float value); | |||
| 123 | FloatAttr getF32FloatAttr(float value); | |||
| 124 | FloatAttr getF64FloatAttr(double value); | |||
| 125 | ||||
| 126 | IntegerAttr getI8IntegerAttr(int8_t value); | |||
| 127 | IntegerAttr getI16IntegerAttr(int16_t value); | |||
| 128 | IntegerAttr getI32IntegerAttr(int32_t value); | |||
| 129 | IntegerAttr getI64IntegerAttr(int64_t value); | |||
| 130 | IntegerAttr getIndexAttr(int64_t value); | |||
| 131 | ||||
| 132 | /// Signed and unsigned integer attribute getters. | |||
| 133 | IntegerAttr getSI32IntegerAttr(int32_t value); | |||
| 134 | IntegerAttr getUI32IntegerAttr(uint32_t value); | |||
| 135 | ||||
| 136 | /// Vector-typed DenseIntElementsAttr getters. `values` must not be empty. | |||
| 137 | DenseIntElementsAttr getBoolVectorAttr(ArrayRef<bool> values); | |||
| 138 | DenseIntElementsAttr getI32VectorAttr(ArrayRef<int32_t> values); | |||
| 139 | DenseIntElementsAttr getI64VectorAttr(ArrayRef<int64_t> values); | |||
| 140 | DenseIntElementsAttr getIndexVectorAttr(ArrayRef<int64_t> values); | |||
| 141 | ||||
| 142 | /// Tensor-typed DenseIntElementsAttr getters. `values` can be empty. | |||
| 143 | /// These are generally preferable for representing general lists of integers | |||
| 144 | /// as attributes. | |||
| 145 | DenseIntElementsAttr getI32TensorAttr(ArrayRef<int32_t> values); | |||
| 146 | DenseIntElementsAttr getI64TensorAttr(ArrayRef<int64_t> values); | |||
| 147 | DenseIntElementsAttr getIndexTensorAttr(ArrayRef<int64_t> values); | |||
| 148 | ||||
| 149 | /// Tensor-typed DenseArrayAttr getters. | |||
| 150 | DenseBoolArrayAttr getDenseBoolArrayAttr(ArrayRef<bool> values); | |||
| 151 | DenseI8ArrayAttr getDenseI8ArrayAttr(ArrayRef<int8_t> values); | |||
| 152 | DenseI16ArrayAttr getDenseI16ArrayAttr(ArrayRef<int16_t> values); | |||
| 153 | DenseI32ArrayAttr getDenseI32ArrayAttr(ArrayRef<int32_t> values); | |||
| 154 | DenseI64ArrayAttr getDenseI64ArrayAttr(ArrayRef<int64_t> values); | |||
| 155 | DenseF32ArrayAttr getDenseF32ArrayAttr(ArrayRef<float> values); | |||
| 156 | DenseF64ArrayAttr getDenseF64ArrayAttr(ArrayRef<double> values); | |||
| 157 | ||||
| 158 | ArrayAttr getAffineMapArrayAttr(ArrayRef<AffineMap> values); | |||
| 159 | ArrayAttr getBoolArrayAttr(ArrayRef<bool> values); | |||
| 160 | ArrayAttr getI32ArrayAttr(ArrayRef<int32_t> values); | |||
| 161 | ArrayAttr getI64ArrayAttr(ArrayRef<int64_t> values); | |||
| 162 | ArrayAttr getIndexArrayAttr(ArrayRef<int64_t> values); | |||
| 163 | ArrayAttr getF32ArrayAttr(ArrayRef<float> values); | |||
| 164 | ArrayAttr getF64ArrayAttr(ArrayRef<double> values); | |||
| 165 | ArrayAttr getStrArrayAttr(ArrayRef<StringRef> values); | |||
| 166 | ArrayAttr getTypeArrayAttr(TypeRange values); | |||
| 167 | ||||
| 168 | // Affine expressions and affine maps. | |||
| 169 | AffineExpr getAffineDimExpr(unsigned position); | |||
| 170 | AffineExpr getAffineSymbolExpr(unsigned position); | |||
| 171 | AffineExpr getAffineConstantExpr(int64_t constant); | |||
| 172 | ||||
| 173 | // Special cases of affine maps and integer sets | |||
| 174 | /// Returns a zero result affine map with no dimensions or symbols: () -> (). | |||
| 175 | AffineMap getEmptyAffineMap(); | |||
| 176 | /// Returns a single constant result affine map with 0 dimensions and 0 | |||
| 177 | /// symbols. One constant result: () -> (val). | |||
| 178 | AffineMap getConstantAffineMap(int64_t val); | |||
| 179 | // One dimension id identity map: (i) -> (i). | |||
| 180 | AffineMap getDimIdentityMap(); | |||
| 181 | // Multi-dimensional identity map: (d0, d1, d2) -> (d0, d1, d2). | |||
| 182 | AffineMap getMultiDimIdentityMap(unsigned rank); | |||
| 183 | // One symbol identity map: ()[s] -> (s). | |||
| 184 | AffineMap getSymbolIdentityMap(); | |||
| 185 | ||||
| 186 | /// Returns a map that shifts its (single) input dimension by 'shift'. | |||
| 187 | /// (d0) -> (d0 + shift) | |||
| 188 | AffineMap getSingleDimShiftAffineMap(int64_t shift); | |||
| 189 | ||||
| 190 | /// Returns an affine map that is a translation (shift) of all result | |||
| 191 | /// expressions in 'map' by 'shift'. | |||
| 192 | /// Eg: input: (d0, d1)[s0] -> (d0, d1 + s0), shift = 2 | |||
| 193 | /// returns: (d0, d1)[s0] -> (d0 + 2, d1 + s0 + 2) | |||
| 194 | AffineMap getShiftedAffineMap(AffineMap map, int64_t shift); | |||
| 195 | ||||
| 196 | protected: | |||
| 197 | MLIRContext *context; | |||
| 198 | }; | |||
| 199 | ||||
| 200 | /// This class helps build Operations. Operations that are created are | |||
| 201 | /// automatically inserted at an insertion point. The builder is copyable. | |||
| 202 | class OpBuilder : public Builder { | |||
| 203 | public: | |||
| 204 | struct Listener; | |||
| 205 | ||||
| 206 | /// Create a builder with the given context. | |||
| 207 | explicit OpBuilder(MLIRContext *ctx, Listener *listener = nullptr) | |||
| 208 | : Builder(ctx), listener(listener) {} | |||
| 209 | ||||
| 210 | /// Create a builder and set the insertion point to the start of the region. | |||
| 211 | explicit OpBuilder(Region *region, Listener *listener = nullptr) | |||
| 212 | : OpBuilder(region->getContext(), listener) { | |||
| 213 | if (!region->empty()) | |||
| 214 | setInsertionPoint(®ion->front(), region->front().begin()); | |||
| 215 | } | |||
| 216 | explicit OpBuilder(Region ®ion, Listener *listener = nullptr) | |||
| 217 | : OpBuilder(®ion, listener) {} | |||
| 218 | ||||
| 219 | /// Create a builder and set insertion point to the given operation, which | |||
| 220 | /// will cause subsequent insertions to go right before it. | |||
| 221 | explicit OpBuilder(Operation *op, Listener *listener = nullptr) | |||
| 222 | : OpBuilder(op->getContext(), listener) { | |||
| 223 | setInsertionPoint(op); | |||
| 224 | } | |||
| 225 | ||||
| 226 | OpBuilder(Block *block, Block::iterator insertPoint, | |||
| 227 | Listener *listener = nullptr) | |||
| 228 | : OpBuilder(block->getParent()->getContext(), listener) { | |||
| 229 | setInsertionPoint(block, insertPoint); | |||
| 230 | } | |||
| 231 | ||||
| 232 | /// Create a builder and set the insertion point to before the first operation | |||
| 233 | /// in the block but still inside the block. | |||
| 234 | static OpBuilder atBlockBegin(Block *block, Listener *listener = nullptr) { | |||
| 235 | return OpBuilder(block, block->begin(), listener); | |||
| 236 | } | |||
| 237 | ||||
| 238 | /// Create a builder and set the insertion point to after the last operation | |||
| 239 | /// in the block but still inside the block. | |||
| 240 | static OpBuilder atBlockEnd(Block *block, Listener *listener = nullptr) { | |||
| 241 | return OpBuilder(block, block->end(), listener); | |||
| 242 | } | |||
| 243 | ||||
| 244 | /// Create a builder and set the insertion point to before the block | |||
| 245 | /// terminator. | |||
| 246 | static OpBuilder atBlockTerminator(Block *block, | |||
| 247 | Listener *listener = nullptr) { | |||
| 248 | auto *terminator = block->getTerminator(); | |||
| 249 | assert(terminator != nullptr && "the block has no terminator")(static_cast <bool> (terminator != nullptr && "the block has no terminator" ) ? void (0) : __assert_fail ("terminator != nullptr && \"the block has no terminator\"" , "llvm/../mlir/include/mlir/IR/Builders.h", 249, __extension__ __PRETTY_FUNCTION__)); | |||
| 250 | return OpBuilder(block, Block::iterator(terminator), listener); | |||
| 251 | } | |||
| 252 | ||||
| 253 | //===--------------------------------------------------------------------===// | |||
| 254 | // Listeners | |||
| 255 | //===--------------------------------------------------------------------===// | |||
| 256 | ||||
| 257 | /// Base class for listeners. | |||
| 258 | struct ListenerBase { | |||
| 259 | /// The kind of listener. | |||
| 260 | enum class Kind { | |||
| 261 | /// OpBuilder::Listener or user-derived class. | |||
| 262 | OpBuilderListener = 0, | |||
| 263 | ||||
| 264 | /// RewriterBase::Listener or user-derived class. | |||
| 265 | RewriterBaseListener = 1 | |||
| 266 | }; | |||
| 267 | ||||
| 268 | Kind getKind() const { return kind; } | |||
| 269 | ||||
| 270 | protected: | |||
| 271 | ListenerBase(Kind kind) : kind(kind) {} | |||
| 272 | ||||
| 273 | private: | |||
| 274 | const Kind kind; | |||
| 275 | }; | |||
| 276 | ||||
| 277 | /// This class represents a listener that may be used to hook into various | |||
| 278 | /// actions within an OpBuilder. | |||
| 279 | struct Listener : public ListenerBase { | |||
| 280 | Listener() : ListenerBase(ListenerBase::Kind::OpBuilderListener) {} | |||
| 281 | ||||
| 282 | virtual ~Listener() = default; | |||
| 283 | ||||
| 284 | /// Notification handler for when an operation is inserted into the builder. | |||
| 285 | /// `op` is the operation that was inserted. | |||
| 286 | virtual void notifyOperationInserted(Operation *op) {} | |||
| 287 | ||||
| 288 | /// Notification handler for when a block is created using the builder. | |||
| 289 | /// `block` is the block that was created. | |||
| 290 | virtual void notifyBlockCreated(Block *block) {} | |||
| 291 | ||||
| 292 | protected: | |||
| 293 | Listener(Kind kind) : ListenerBase(kind) {} | |||
| 294 | }; | |||
| 295 | ||||
| 296 | /// Sets the listener of this builder to the one provided. | |||
| 297 | void setListener(Listener *newListener) { listener = newListener; } | |||
| 298 | ||||
| 299 | /// Returns the current listener of this builder, or nullptr if this builder | |||
| 300 | /// doesn't have a listener. | |||
| 301 | Listener *getListener() const { return listener; } | |||
| 302 | ||||
| 303 | //===--------------------------------------------------------------------===// | |||
| 304 | // Insertion Point Management | |||
| 305 | //===--------------------------------------------------------------------===// | |||
| 306 | ||||
| 307 | /// This class represents a saved insertion point. | |||
| 308 | class InsertPoint { | |||
| 309 | public: | |||
| 310 | /// Creates a new insertion point which doesn't point to anything. | |||
| 311 | InsertPoint() = default; | |||
| 312 | ||||
| 313 | /// Creates a new insertion point at the given location. | |||
| 314 | InsertPoint(Block *insertBlock, Block::iterator insertPt) | |||
| 315 | : block(insertBlock), point(insertPt) {} | |||
| 316 | ||||
| 317 | /// Returns true if this insert point is set. | |||
| 318 | bool isSet() const { return (block != nullptr); } | |||
| 319 | ||||
| 320 | Block *getBlock() const { return block; } | |||
| 321 | Block::iterator getPoint() const { return point; } | |||
| 322 | ||||
| 323 | private: | |||
| 324 | Block *block = nullptr; | |||
| 325 | Block::iterator point; | |||
| 326 | }; | |||
| 327 | ||||
| 328 | /// RAII guard to reset the insertion point of the builder when destroyed. | |||
| 329 | class InsertionGuard { | |||
| 330 | public: | |||
| 331 | InsertionGuard(OpBuilder &builder) | |||
| 332 | : builder(&builder), ip(builder.saveInsertionPoint()) {} | |||
| 333 | ||||
| 334 | ~InsertionGuard() { | |||
| 335 | if (builder) | |||
| 336 | builder->restoreInsertionPoint(ip); | |||
| 337 | } | |||
| 338 | ||||
| 339 | InsertionGuard(const InsertionGuard &) = delete; | |||
| 340 | InsertionGuard &operator=(const InsertionGuard &) = delete; | |||
| 341 | ||||
| 342 | /// Implement the move constructor to clear the builder field of `other`. | |||
| 343 | /// That way it does not restore the insertion point upon destruction as | |||
| 344 | /// that should be done exclusively by the just constructed InsertionGuard. | |||
| 345 | InsertionGuard(InsertionGuard &&other) noexcept | |||
| 346 | : builder(other.builder), ip(other.ip) { | |||
| 347 | other.builder = nullptr; | |||
| 348 | } | |||
| 349 | ||||
| 350 | InsertionGuard &operator=(InsertionGuard &&other) = delete; | |||
| 351 | ||||
| 352 | private: | |||
| 353 | OpBuilder *builder; | |||
| 354 | OpBuilder::InsertPoint ip; | |||
| 355 | }; | |||
| 356 | ||||
| 357 | /// Reset the insertion point to no location. Creating an operation without a | |||
| 358 | /// set insertion point is an error, but this can still be useful when the | |||
| 359 | /// current insertion point a builder refers to is being removed. | |||
| 360 | void clearInsertionPoint() { | |||
| 361 | this->block = nullptr; | |||
| 362 | insertPoint = Block::iterator(); | |||
| 363 | } | |||
| 364 | ||||
| 365 | /// Return a saved insertion point. | |||
| 366 | InsertPoint saveInsertionPoint() const { | |||
| 367 | return InsertPoint(getInsertionBlock(), getInsertionPoint()); | |||
| 368 | } | |||
| 369 | ||||
| 370 | /// Restore the insert point to a previously saved point. | |||
| 371 | void restoreInsertionPoint(InsertPoint ip) { | |||
| 372 | if (ip.isSet()) | |||
| 373 | setInsertionPoint(ip.getBlock(), ip.getPoint()); | |||
| 374 | else | |||
| 375 | clearInsertionPoint(); | |||
| 376 | } | |||
| 377 | ||||
| 378 | /// Set the insertion point to the specified location. | |||
| 379 | void setInsertionPoint(Block *block, Block::iterator insertPoint) { | |||
| 380 | // TODO: check that insertPoint is in this rather than some other block. | |||
| 381 | this->block = block; | |||
| 382 | this->insertPoint = insertPoint; | |||
| 383 | } | |||
| 384 | ||||
| 385 | /// Sets the insertion point to the specified operation, which will cause | |||
| 386 | /// subsequent insertions to go right before it. | |||
| 387 | void setInsertionPoint(Operation *op) { | |||
| 388 | setInsertionPoint(op->getBlock(), Block::iterator(op)); | |||
| 389 | } | |||
| 390 | ||||
| 391 | /// Sets the insertion point to the node after the specified operation, which | |||
| 392 | /// will cause subsequent insertions to go right after it. | |||
| 393 | void setInsertionPointAfter(Operation *op) { | |||
| 394 | setInsertionPoint(op->getBlock(), ++Block::iterator(op)); | |||
| 395 | } | |||
| 396 | ||||
| 397 | /// Sets the insertion point to the node after the specified value. If value | |||
| 398 | /// has a defining operation, sets the insertion point to the node after such | |||
| 399 | /// defining operation. This will cause subsequent insertions to go right | |||
| 400 | /// after it. Otherwise, value is a BlockArgument. Sets the insertion point to | |||
| 401 | /// the start of its block. | |||
| 402 | void setInsertionPointAfterValue(Value val) { | |||
| 403 | if (Operation *op = val.getDefiningOp()) { | |||
| 404 | setInsertionPointAfter(op); | |||
| 405 | } else { | |||
| 406 | auto blockArg = val.cast<BlockArgument>(); | |||
| 407 | setInsertionPointToStart(blockArg.getOwner()); | |||
| 408 | } | |||
| 409 | } | |||
| 410 | ||||
| 411 | /// Sets the insertion point to the start of the specified block. | |||
| 412 | void setInsertionPointToStart(Block *block) { | |||
| 413 | setInsertionPoint(block, block->begin()); | |||
| 414 | } | |||
| 415 | ||||
| 416 | /// Sets the insertion point to the end of the specified block. | |||
| 417 | void setInsertionPointToEnd(Block *block) { | |||
| 418 | setInsertionPoint(block, block->end()); | |||
| 419 | } | |||
| 420 | ||||
| 421 | /// Return the block the current insertion point belongs to. Note that the | |||
| 422 | /// insertion point is not necessarily the end of the block. | |||
| 423 | Block *getInsertionBlock() const { return block; } | |||
| 424 | ||||
| 425 | /// Returns the current insertion point of the builder. | |||
| 426 | Block::iterator getInsertionPoint() const { return insertPoint; } | |||
| 427 | ||||
| 428 | /// Returns the current block of the builder. | |||
| 429 | Block *getBlock() const { return block; } | |||
| 430 | ||||
| 431 | //===--------------------------------------------------------------------===// | |||
| 432 | // Block Creation | |||
| 433 | //===--------------------------------------------------------------------===// | |||
| 434 | ||||
| 435 | /// Add new block with 'argTypes' arguments and set the insertion point to the | |||
| 436 | /// end of it. The block is inserted at the provided insertion point of | |||
| 437 | /// 'parent'. `locs` contains the locations of the inserted arguments, and | |||
| 438 | /// should match the size of `argTypes`. | |||
| 439 | Block *createBlock(Region *parent, Region::iterator insertPt = {}, | |||
| 440 | TypeRange argTypes = std::nullopt, | |||
| 441 | ArrayRef<Location> locs = std::nullopt); | |||
| 442 | ||||
| 443 | /// Add new block with 'argTypes' arguments and set the insertion point to the | |||
| 444 | /// end of it. The block is placed before 'insertBefore'. `locs` contains the | |||
| 445 | /// locations of the inserted arguments, and should match the size of | |||
| 446 | /// `argTypes`. | |||
| 447 | Block *createBlock(Block *insertBefore, TypeRange argTypes = std::nullopt, | |||
| 448 | ArrayRef<Location> locs = std::nullopt); | |||
| 449 | ||||
| 450 | //===--------------------------------------------------------------------===// | |||
| 451 | // Operation Creation | |||
| 452 | //===--------------------------------------------------------------------===// | |||
| 453 | ||||
| 454 | /// Insert the given operation at the current insertion point and return it. | |||
| 455 | Operation *insert(Operation *op); | |||
| 456 | ||||
| 457 | /// Creates an operation given the fields represented as an OperationState. | |||
| 458 | Operation *create(const OperationState &state); | |||
| 459 | ||||
| 460 | /// Creates an operation with the given fields. | |||
| 461 | Operation *create(Location loc, StringAttr opName, ValueRange operands, | |||
| 462 | TypeRange types = {}, | |||
| 463 | ArrayRef<NamedAttribute> attributes = {}, | |||
| 464 | BlockRange successors = {}, | |||
| 465 | MutableArrayRef<std::unique_ptr<Region>> regions = {}); | |||
| 466 | ||||
| 467 | private: | |||
| 468 | /// Helper for sanity checking preconditions for create* methods below. | |||
| 469 | template <typename OpT> | |||
| 470 | RegisteredOperationName getCheckRegisteredInfo(MLIRContext *ctx) { | |||
| 471 | std::optional<RegisteredOperationName> opName = | |||
| 472 | RegisteredOperationName::lookup(OpT::getOperationName(), ctx); | |||
| 473 | if (LLVM_UNLIKELY(!opName)__builtin_expect((bool)(!opName), false)) { | |||
| 474 | llvm::report_fatal_error( | |||
| 475 | "Building op `" + OpT::getOperationName() + | |||
| 476 | "` but it isn't registered in this MLIRContext: the dialect may not " | |||
| 477 | "be loaded or this operation isn't registered by the dialect. See " | |||
| 478 | "also https://mlir.llvm.org/getting_started/Faq/" | |||
| 479 | "#registered-loaded-dependent-whats-up-with-dialects-management"); | |||
| 480 | } | |||
| 481 | return *opName; | |||
| 482 | } | |||
| 483 | ||||
| 484 | public: | |||
| 485 | /// Create an operation of specific op type at the current insertion point. | |||
| 486 | template <typename OpTy, typename... Args> | |||
| 487 | OpTy create(Location location, Args &&...args) { | |||
| 488 | OperationState state(location, | |||
| 489 | getCheckRegisteredInfo<OpTy>(location.getContext())); | |||
| 490 | OpTy::build(*this, state, std::forward<Args>(args)...); | |||
| ||||
| 491 | auto *op = create(state); | |||
| 492 | auto result = dyn_cast<OpTy>(op); | |||
| 493 | assert(result && "builder didn't return the right type")(static_cast <bool> (result && "builder didn't return the right type" ) ? void (0) : __assert_fail ("result && \"builder didn't return the right type\"" , "llvm/../mlir/include/mlir/IR/Builders.h", 493, __extension__ __PRETTY_FUNCTION__)); | |||
| 494 | return result; | |||
| 495 | } | |||
| 496 | ||||
| 497 | /// Create an operation of specific op type at the current insertion point, | |||
| 498 | /// and immediately try to fold it. This functions populates 'results' with | |||
| 499 | /// the results after folding the operation. | |||
| 500 | template <typename OpTy, typename... Args> | |||
| 501 | void createOrFold(SmallVectorImpl<Value> &results, Location location, | |||
| 502 | Args &&...args) { | |||
| 503 | // Create the operation without using 'create' as we don't want to | |||
| 504 | // insert it yet. | |||
| 505 | OperationState state(location, | |||
| 506 | getCheckRegisteredInfo<OpTy>(location.getContext())); | |||
| 507 | OpTy::build(*this, state, std::forward<Args>(args)...); | |||
| 508 | Operation *op = Operation::create(state); | |||
| 509 | ||||
| 510 | // Fold the operation. If successful destroy it, otherwise insert it. | |||
| 511 | if (succeeded(tryFold(op, results))) | |||
| 512 | op->destroy(); | |||
| 513 | else | |||
| 514 | insert(op); | |||
| 515 | } | |||
| 516 | ||||
| 517 | /// Overload to create or fold a single result operation. | |||
| 518 | template <typename OpTy, typename... Args> | |||
| 519 | std::enable_if_t<OpTy::template hasTrait<OpTrait::OneResult>(), Value> | |||
| 520 | createOrFold(Location location, Args &&...args) { | |||
| 521 | SmallVector<Value, 1> results; | |||
| 522 | createOrFold<OpTy>(results, location, std::forward<Args>(args)...); | |||
| 523 | return results.front(); | |||
| 524 | } | |||
| 525 | ||||
| 526 | /// Overload to create or fold a zero result operation. | |||
| 527 | template <typename OpTy, typename... Args> | |||
| 528 | std::enable_if_t<OpTy::template hasTrait<OpTrait::ZeroResults>(), OpTy> | |||
| 529 | createOrFold(Location location, Args &&...args) { | |||
| 530 | auto op = create<OpTy>(location, std::forward<Args>(args)...); | |||
| 531 | SmallVector<Value, 0> unused; | |||
| 532 | (void)tryFold(op.getOperation(), unused); | |||
| 533 | ||||
| 534 | // Folding cannot remove a zero-result operation, so for convenience we | |||
| 535 | // continue to return it. | |||
| 536 | return op; | |||
| 537 | } | |||
| 538 | ||||
| 539 | /// Attempts to fold the given operation and places new results within | |||
| 540 | /// 'results'. Returns success if the operation was folded, failure otherwise. | |||
| 541 | /// Note: This function does not erase the operation on a successful fold. | |||
| 542 | LogicalResult tryFold(Operation *op, SmallVectorImpl<Value> &results); | |||
| 543 | ||||
| 544 | /// Creates a deep copy of the specified operation, remapping any operands | |||
| 545 | /// that use values outside of the operation using the map that is provided | |||
| 546 | /// ( leaving them alone if no entry is present). Replaces references to | |||
| 547 | /// cloned sub-operations to the corresponding operation that is copied, | |||
| 548 | /// and adds those mappings to the map. | |||
| 549 | Operation *clone(Operation &op, IRMapping &mapper); | |||
| 550 | Operation *clone(Operation &op); | |||
| 551 | ||||
| 552 | /// Creates a deep copy of this operation but keep the operation regions | |||
| 553 | /// empty. Operands are remapped using `mapper` (if present), and `mapper` is | |||
| 554 | /// updated to contain the results. | |||
| 555 | Operation *cloneWithoutRegions(Operation &op, IRMapping &mapper) { | |||
| 556 | return insert(op.cloneWithoutRegions(mapper)); | |||
| 557 | } | |||
| 558 | Operation *cloneWithoutRegions(Operation &op) { | |||
| 559 | return insert(op.cloneWithoutRegions()); | |||
| 560 | } | |||
| 561 | template <typename OpT> | |||
| 562 | OpT cloneWithoutRegions(OpT op) { | |||
| 563 | return cast<OpT>(cloneWithoutRegions(*op.getOperation())); | |||
| 564 | } | |||
| 565 | ||||
| 566 | protected: | |||
| 567 | /// The optional listener for events of this builder. | |||
| 568 | Listener *listener; | |||
| 569 | ||||
| 570 | private: | |||
| 571 | /// The current block this builder is inserting into. | |||
| 572 | Block *block = nullptr; | |||
| 573 | /// The insertion point within the block that this builder is inserting | |||
| 574 | /// before. | |||
| 575 | Block::iterator insertPoint; | |||
| 576 | }; | |||
| 577 | ||||
| 578 | } // namespace mlir | |||
| 579 | ||||
| 580 | #endif |