| 1 | //====- LowerToLLVM.cpp - Lowering from CIR to LLVMIR ---------------------===// |
| 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 | // This file implements lowering of CIR operations to LLVMIR. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "LowerToLLVM.h" |
| 14 | #include "LowerToLLVMMetadata.h" |
| 15 | #include "LowerToLLVMOffloadPrintf.h" |
| 16 | #include "LowerToLLVMOpenCLMetadata.h" |
| 17 | |
| 18 | #include <array> |
| 19 | #include <optional> |
| 20 | |
| 21 | #include "mlir/Conversion/LLVMCommon/TypeConverter.h" |
| 22 | #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h" |
| 23 | #include "mlir/Dialect/DLTI/DLTI.h" |
| 24 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 25 | #include "mlir/Dialect/LLVMIR/LLVMAttrs.h" |
| 26 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" |
| 27 | #include "mlir/Dialect/LLVMIR/LLVMTypes.h" |
| 28 | #include "mlir/Dialect/LLVMIR/Transforms/Passes.h" |
| 29 | #include "mlir/Dialect/OpenMP/OpenMPDialect.h" |
| 30 | #include "mlir/Dialect/OpenMP/Transforms/Passes.h" |
| 31 | #include "mlir/Dialect/Ptr/IR/MemorySpaceInterfaces.h" |
| 32 | #include "mlir/IR/BuiltinAttributes.h" |
| 33 | #include "mlir/IR/BuiltinDialect.h" |
| 34 | #include "mlir/IR/BuiltinOps.h" |
| 35 | #include "mlir/IR/Diagnostics.h" |
| 36 | #include "mlir/IR/Location.h" |
| 37 | #include "mlir/IR/Matchers.h" |
| 38 | #include "mlir/IR/Types.h" |
| 39 | #include "mlir/Pass/Pass.h" |
| 40 | #include "mlir/Pass/PassManager.h" |
| 41 | #include "mlir/Support/LLVM.h" |
| 42 | #include "mlir/Target/LLVMIR/Dialect/Builtin/BuiltinToLLVMIRTranslation.h" |
| 43 | #include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h" |
| 44 | #include "mlir/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.h" |
| 45 | #include "mlir/Target/LLVMIR/Export.h" |
| 46 | #include "mlir/Transforms/DialectConversion.h" |
| 47 | #include "clang/CIR/Dialect/IR/CIRAttrs.h" |
| 48 | #include "clang/CIR/Dialect/IR/CIRDataLayout.h" |
| 49 | #include "clang/CIR/Dialect/IR/CIRDialect.h" |
| 50 | #include "clang/CIR/Dialect/IR/CIRTypes.h" |
| 51 | #include "clang/CIR/Dialect/Passes.h" |
| 52 | #include "clang/CIR/Dialect/Transforms/CIRTransformUtils.h" |
| 53 | #include "clang/CIR/LoweringHelpers.h" |
| 54 | #include "clang/CIR/MissingFeatures.h" |
| 55 | #include "clang/CIR/Passes.h" |
| 56 | #include "llvm/ADT/MapVector.h" |
| 57 | #include "llvm/ADT/StringMap.h" |
| 58 | #include "llvm/ADT/TypeSwitch.h" |
| 59 | #include "llvm/IR/Module.h" |
| 60 | #include "llvm/Support/AMDGPUAddrSpace.h" |
| 61 | #include "llvm/Support/Casting.h" |
| 62 | #include "llvm/Support/ErrorHandling.h" |
| 63 | #include "llvm/Support/TimeProfiler.h" |
| 64 | #include "llvm/Support/VirtualFileSystem.h" |
| 65 | #include "llvm/Support/raw_ostream.h" |
| 66 | #include "llvm/TargetParser/AtomicScope.h" |
| 67 | |
| 68 | using namespace cir; |
| 69 | using namespace llvm; |
| 70 | |
| 71 | namespace cir { |
| 72 | namespace direct { |
| 73 | |
| 74 | //===----------------------------------------------------------------------===// |
| 75 | // Helper Methods |
| 76 | //===----------------------------------------------------------------------===// |
| 77 | |
| 78 | namespace { |
| 79 | /// If the given type is a vector or matrix type, return its element type. |
| 80 | /// Otherwise return the given type unchanged. |
| 81 | mlir::Type elementTypeIfVectorOrMatrix(mlir::Type type) { |
| 82 | return llvm::TypeSwitch<mlir::Type, mlir::Type>(type) |
| 83 | .Case<cir::VectorType, cir::MatrixType, mlir::VectorType>( |
| 84 | caseFn: [](auto p) { return p.getElementType(); }) |
| 85 | .Default(defaultFn: [](mlir::Type p) { return p; }); |
| 86 | } |
| 87 | } // namespace |
| 88 | |
| 89 | static constexpr char kTypeConversionErrorMsg[] = |
| 90 | "ClangIR lowering NYI: cannot convert " ; |
| 91 | |
| 92 | static mlir::InFlightDiagnostic emitTypeConversionError(mlir::Operation *op, |
| 93 | mlir::Type type) { |
| 94 | return op->emitError() << kTypeConversionErrorMsg << type; |
| 95 | } |
| 96 | |
| 97 | static mlir::InFlightDiagnostic emitTypeConversionError(mlir::Location loc, |
| 98 | mlir::Type type) { |
| 99 | return mlir::emitError(loc) << kTypeConversionErrorMsg << type; |
| 100 | } |
| 101 | |
| 102 | /// Alignment to use for a memory access whose op carries no explicit alignment. |
| 103 | /// For _BitInt the storage integer iM's ABI alignment (e.g. i128's 16) |
| 104 | /// over-aligns the value, so use the _BitInt language alignment (e.g. 8). |
| 105 | static uint64_t getMemoryFallbackAlignment(mlir::Operation *op, |
| 106 | mlir::Type cirType, |
| 107 | mlir::Type llvmMemType, |
| 108 | const mlir::DataLayout &dataLayout) { |
| 109 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val&: cirType); |
| 110 | intTy && intTy.isBitInt()) |
| 111 | return cir::CIRDataLayout(op->getParentOfType<mlir::ModuleOp>()) |
| 112 | .getTypeLanguageAlignment(ty: intTy); |
| 113 | return dataLayout.getTypeABIAlignment(t: llvmMemType); |
| 114 | } |
| 115 | |
| 116 | static mlir::Value createIntCast(mlir::OpBuilder &bld, mlir::Value src, |
| 117 | mlir::IntegerType dstTy, |
| 118 | bool isSigned = false) { |
| 119 | mlir::Type srcTy = src.getType(); |
| 120 | assert(mlir::isa<mlir::IntegerType>(srcTy)); |
| 121 | |
| 122 | unsigned srcWidth = mlir::cast<mlir::IntegerType>(Val&: srcTy).getWidth(); |
| 123 | unsigned dstWidth = mlir::cast<mlir::IntegerType>(Val&: dstTy).getWidth(); |
| 124 | mlir::Location loc = src.getLoc(); |
| 125 | |
| 126 | if (dstWidth > srcWidth && isSigned) |
| 127 | return mlir::LLVM::SExtOp::create(builder&: bld, location: loc, res: dstTy, arg: src); |
| 128 | if (dstWidth > srcWidth) |
| 129 | return mlir::LLVM::ZExtOp::create(builder&: bld, location: loc, res: dstTy, arg: src); |
| 130 | if (dstWidth < srcWidth) |
| 131 | return mlir::LLVM::TruncOp::create(builder&: bld, location: loc, res: dstTy, arg: src); |
| 132 | return mlir::LLVM::BitcastOp::create(builder&: bld, location: loc, res: dstTy, arg: src); |
| 133 | } |
| 134 | |
| 135 | /// Cast a _BitInt(N) value between its literal width iN and its padded |
| 136 | /// in-memory storage iM (sign/zero-extend to memory per signedness so the |
| 137 | /// padding bits are well-defined, truncate back on load; matches classic |
| 138 | /// CodeGen). Callers must first rule out the unsupported byte-array storage |
| 139 | /// form (isSplitStorageBitInt / a null convertTypeForMemory result). |
| 140 | static mlir::Value |
| 141 | castBitIntMemoryStorage(mlir::ConversionPatternRewriter &rewriter, |
| 142 | const mlir::DataLayout &dataLayout, cir::IntType intTy, |
| 143 | mlir::Value value, bool toMemory) { |
| 144 | unsigned storageBits = intTy.getStorageTypeWidth(dataLayout); |
| 145 | if (storageBits == intTy.getWidth()) |
| 146 | return value; |
| 147 | unsigned dstBits = toMemory ? storageBits : intTy.getWidth(); |
| 148 | return createIntCast(bld&: rewriter, src: value, dstTy: rewriter.getIntegerType(width: dstBits), |
| 149 | /*isSigned=*/toMemory && intTy.isSigned()); |
| 150 | } |
| 151 | |
| 152 | static mlir::LLVM::Visibility |
| 153 | lowerCIRVisibilityToLLVMVisibility(cir::VisibilityKind visibilityKind) { |
| 154 | switch (visibilityKind) { |
| 155 | case cir::VisibilityKind::Default: |
| 156 | return ::mlir::LLVM::Visibility::Default; |
| 157 | case cir::VisibilityKind::Hidden: |
| 158 | return ::mlir::LLVM::Visibility::Hidden; |
| 159 | case cir::VisibilityKind::Protected: |
| 160 | return ::mlir::LLVM::Visibility::Protected; |
| 161 | } |
| 162 | llvm_unreachable("fully covered switch" ); |
| 163 | } |
| 164 | |
| 165 | static mlir::Value |
| 166 | emitBoolVecConversion(mlir::ConversionPatternRewriter &rewriter, |
| 167 | mlir::Value srcVec, unsigned numElementsDst) { |
| 168 | auto srcTy = mlir::cast<mlir::VectorType>(Val: srcVec.getType()); |
| 169 | unsigned numElementsSrc = srcTy.getNumElements(); |
| 170 | if (numElementsSrc == numElementsDst) |
| 171 | return srcVec; |
| 172 | |
| 173 | SmallVector<int32_t, 8> mask(numElementsDst, -1); |
| 174 | for (unsigned i : llvm::seq(Size: std::min(a: numElementsDst, b: numElementsSrc))) |
| 175 | mask[i] = i; |
| 176 | |
| 177 | mlir::Location loc = srcVec.getLoc(); |
| 178 | auto poison = mlir::LLVM::PoisonOp::create(builder&: rewriter, location: loc, res: srcVec.getType()); |
| 179 | return mlir::LLVM::ShuffleVectorOp::create(builder&: rewriter, location: loc, v1: srcVec, v2: poison, |
| 180 | mask); |
| 181 | } |
| 182 | |
| 183 | /// Emits the value from memory as expected by its users. Should be called when |
| 184 | /// the memory represetnation of a CIR type is not equal to its scalar |
| 185 | /// representation. |
| 186 | static mlir::Value emitFromMemory(mlir::ConversionPatternRewriter &rewriter, |
| 187 | const mlir::TypeConverter &converter, |
| 188 | mlir::DataLayout const &dataLayout, |
| 189 | cir::LoadOp op, mlir::Value value) { |
| 190 | // TODO(cir): Handle other types similarly to clang's codegen EmitFromMemory |
| 191 | if (auto boolTy = mlir::dyn_cast<cir::BoolType>(Val: op.getType())) { |
| 192 | // Create a cast value from specified size in datalayout to i1 |
| 193 | assert(value.getType().isInteger(dataLayout.getTypeSizeInBits(boolTy))); |
| 194 | return createIntCast(bld&: rewriter, src: value, dstTy: rewriter.getI1Type()); |
| 195 | } |
| 196 | |
| 197 | // Convert the `iN` back to boolean vectors |
| 198 | if (auto vecTy = mlir::dyn_cast<cir::VectorType>(Val: op.getType())) { |
| 199 | if (mlir::isa<cir::BoolType>(Val: vecTy.getElementType())) { |
| 200 | if (vecTy.isScalable()) |
| 201 | return value; |
| 202 | auto rawIntTy = mlir::cast<mlir::IntegerType>(Val: value.getType()); |
| 203 | auto paddedVecTy = cir::VectorType::getFixed(elementType: vecTy.getElementType(), |
| 204 | size: rawIntTy.getWidth()); |
| 205 | mlir::Type mlirVecTy = converter.convertType(t: paddedVecTy); |
| 206 | // Bitcast iP --> <P x i1>. |
| 207 | auto v = mlir::LLVM::BitcastOp::create(builder&: rewriter, location: value.getLoc(), |
| 208 | res: mlirVecTy, arg: value); |
| 209 | // Shuffle <P x i1> --> <N x i1> (N is the actual bit size). |
| 210 | return emitBoolVecConversion(rewriter, srcVec: v, numElementsDst: vecTy.getFixedSize()); |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | // Truncate the padded storage integer back to the _BitInt's literal width. |
| 215 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val: op.getType()); |
| 216 | intTy && intTy.isBitInt()) |
| 217 | return castBitIntMemoryStorage(rewriter, dataLayout, intTy, value, |
| 218 | /*toMemory=*/false); |
| 219 | |
| 220 | return value; |
| 221 | } |
| 222 | |
| 223 | /// Emits a value to memory with the expected scalar type. Should be called when |
| 224 | /// the memory represetnation of a CIR type is not equal to its scalar |
| 225 | /// representation. |
| 226 | static mlir::Value emitToMemory(mlir::ConversionPatternRewriter &rewriter, |
| 227 | mlir::DataLayout const &dataLayout, |
| 228 | mlir::Type origType, mlir::Value value) { |
| 229 | |
| 230 | // TODO(cir): Handle other types similarly to clang's codegen EmitToMemory |
| 231 | if (auto boolTy = mlir::dyn_cast<cir::BoolType>(Val&: origType)) { |
| 232 | // Create zext of value from i1 to i8 |
| 233 | mlir::IntegerType memType = |
| 234 | rewriter.getIntegerType(width: dataLayout.getTypeSizeInBits(t: boolTy)); |
| 235 | return createIntCast(bld&: rewriter, src: value, dstTy: memType); |
| 236 | } |
| 237 | |
| 238 | // Boolean vectors use: |
| 239 | // * `iN` for fixed-width vectors, |
| 240 | // * `<vscale x N x i1>` for scalable vectors, |
| 241 | // as storage type. |
| 242 | if (auto vecTy = mlir::dyn_cast<cir::VectorType>(Val&: origType)) { |
| 243 | if (mlir::isa<cir::BoolType>(Val: vecTy.getElementType())) { |
| 244 | mlir::Type resultTy; |
| 245 | if (vecTy.isScalable()) |
| 246 | resultTy = |
| 247 | mlir::VectorType::get(shape: vecTy.getSize().getKnownMinValue(), |
| 248 | elementType: vecTy.getElementType(), scalableDims: vecTy.isScalable()); |
| 249 | else { |
| 250 | resultTy = mlir::IntegerType::get(context: origType.getContext(), |
| 251 | width: vecTy.getBoolStorageWidth()); |
| 252 | value = |
| 253 | emitBoolVecConversion(rewriter, srcVec: value, numElementsDst: vecTy.getBoolStorageWidth()); |
| 254 | } |
| 255 | return mlir::LLVM::BitcastOp::create(builder&: rewriter, location: value.getLoc(), res: resultTy, |
| 256 | arg: value); |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | // Sign/zero-extend the _BitInt value to its padded storage integer. |
| 261 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val&: origType); |
| 262 | intTy && intTy.isBitInt()) |
| 263 | return castBitIntMemoryStorage(rewriter, dataLayout, intTy, value, |
| 264 | /*toMemory=*/true); |
| 265 | |
| 266 | return value; |
| 267 | } |
| 268 | |
| 269 | mlir::LLVM::Linkage convertLinkage(cir::GlobalLinkageKind linkage) { |
| 270 | using CIR = cir::GlobalLinkageKind; |
| 271 | using LLVM = mlir::LLVM::Linkage; |
| 272 | |
| 273 | switch (linkage) { |
| 274 | case CIR::AppendingLinkage: |
| 275 | return LLVM::Appending; |
| 276 | case CIR::AvailableExternallyLinkage: |
| 277 | return LLVM::AvailableExternally; |
| 278 | case CIR::CommonLinkage: |
| 279 | return LLVM::Common; |
| 280 | case CIR::ExternalLinkage: |
| 281 | return LLVM::External; |
| 282 | case CIR::ExternalWeakLinkage: |
| 283 | return LLVM::ExternWeak; |
| 284 | case CIR::InternalLinkage: |
| 285 | return LLVM::Internal; |
| 286 | case CIR::LinkOnceAnyLinkage: |
| 287 | return LLVM::Linkonce; |
| 288 | case CIR::LinkOnceODRLinkage: |
| 289 | return LLVM::LinkonceODR; |
| 290 | case CIR::PrivateLinkage: |
| 291 | return LLVM::Private; |
| 292 | case CIR::WeakAnyLinkage: |
| 293 | return LLVM::Weak; |
| 294 | case CIR::WeakODRLinkage: |
| 295 | return LLVM::WeakODR; |
| 296 | }; |
| 297 | llvm_unreachable("Unknown CIR linkage type" ); |
| 298 | } |
| 299 | |
| 300 | static mlir::LLVM::CConv convertCallingConv(cir::CallingConv callingConv) { |
| 301 | using CIR = cir::CallingConv; |
| 302 | using LLVM = mlir::LLVM::CConv; |
| 303 | |
| 304 | switch (callingConv) { |
| 305 | case CIR::C: |
| 306 | return LLVM::C; |
| 307 | case CIR::SpirKernel: |
| 308 | return LLVM::SPIR_KERNEL; |
| 309 | case CIR::SpirFunction: |
| 310 | return LLVM::SPIR_FUNC; |
| 311 | case CIR::PTXKernel: |
| 312 | return LLVM::PTX_Kernel; |
| 313 | case CIR::AMDGPUKernel: |
| 314 | return LLVM::AMDGPU_KERNEL; |
| 315 | } |
| 316 | llvm_unreachable("Unknown calling convention" ); |
| 317 | } |
| 318 | |
| 319 | static mlir::LLVM::uwtable::UWTableKind |
| 320 | convertUWTableKind(cir::UnwindTableKind kind) { |
| 321 | using CIR = cir::UnwindTableKind; |
| 322 | using LLVM = mlir::LLVM::uwtable::UWTableKind; |
| 323 | |
| 324 | switch (kind) { |
| 325 | case CIR::None: |
| 326 | return LLVM::None; |
| 327 | case CIR::Sync: |
| 328 | return LLVM::Sync; |
| 329 | case CIR::Async: |
| 330 | return LLVM::Async; |
| 331 | } |
| 332 | llvm_unreachable("Unknown CIR unwind table kind" ); |
| 333 | } |
| 334 | |
| 335 | static mlir::LLVM::framePointerKind::FramePointerKind |
| 336 | convertFramePointerKind(cir::FramePointerKind kind) { |
| 337 | using CIR = cir::FramePointerKind; |
| 338 | using LLVM = mlir::LLVM::framePointerKind::FramePointerKind; |
| 339 | |
| 340 | switch (kind) { |
| 341 | case CIR::None: |
| 342 | return LLVM::None; |
| 343 | case CIR::NonLeaf: |
| 344 | return LLVM::NonLeaf; |
| 345 | case CIR::All: |
| 346 | return LLVM::All; |
| 347 | case CIR::Reserved: |
| 348 | return LLVM::Reserved; |
| 349 | case CIR::NonLeafNoReserve: |
| 350 | return LLVM::NonLeafNoReserve; |
| 351 | } |
| 352 | llvm_unreachable("Unknown CIR frame pointer kind" ); |
| 353 | } |
| 354 | |
| 355 | /// The "frame-pointer" module flag holds an llvm::FramePointerKind value, as |
| 356 | /// written by llvm::Module::setFramePointer. |
| 357 | static llvm::FramePointerKind |
| 358 | convertFramePointerKindToLLVMIR(cir::FramePointerKind kind) { |
| 359 | using CIR = cir::FramePointerKind; |
| 360 | using LLVM = llvm::FramePointerKind; |
| 361 | |
| 362 | switch (kind) { |
| 363 | case CIR::None: |
| 364 | return LLVM::None; |
| 365 | case CIR::NonLeaf: |
| 366 | return LLVM::NonLeaf; |
| 367 | case CIR::All: |
| 368 | return LLVM::All; |
| 369 | case CIR::Reserved: |
| 370 | return LLVM::Reserved; |
| 371 | case CIR::NonLeafNoReserve: |
| 372 | return LLVM::NonLeafNoReserve; |
| 373 | } |
| 374 | llvm_unreachable("Unknown CIR frame pointer kind" ); |
| 375 | } |
| 376 | |
| 377 | mlir::LogicalResult CIRToLLVMCopyOpLowering::matchAndRewrite( |
| 378 | cir::CopyOp op, OpAdaptor adaptor, |
| 379 | mlir::ConversionPatternRewriter &rewriter) const { |
| 380 | // Opaque pointer conversion does not validate the pointee's memory layout. |
| 381 | mlir::Type pointeeType = op.getType().getPointee(); |
| 382 | if (!convertTypeForMemory(converter: *getTypeConverter(), dataLayout, type: pointeeType)) |
| 383 | return emitTypeConversionError(op, type: pointeeType); |
| 384 | |
| 385 | cir::CIRDataLayout cirDataLayout(op->getParentOfType<mlir::ModuleOp>()); |
| 386 | const mlir::Value length = mlir::LLVM::ConstantOp::create( |
| 387 | builder&: rewriter, location: op.getLoc(), type: rewriter.getI64Type(), |
| 388 | value: op.getCopySizeInBytes(dl: cirDataLayout.layout)); |
| 389 | assert(!cir::MissingFeatures::aggValueSlotVolatile()); |
| 390 | |
| 391 | uint64_t dstTypeAlign = cirDataLayout.getTypeLanguageAlignment( |
| 392 | ty: op.getDst().getType().getPointee()); |
| 393 | uint64_t srcTypeAlign = cirDataLayout.getTypeLanguageAlignment( |
| 394 | ty: op.getSrc().getType().getPointee()); |
| 395 | |
| 396 | mlir::NamedAttribute dstAlignAttr = rewriter.getNamedAttr( |
| 397 | name: mlir::LLVM::LLVMDialect::getAlignAttrName(), |
| 398 | val: rewriter.getI64IntegerAttr(value: op.getDstAlignment().value_or(u&: dstTypeAlign))); |
| 399 | mlir::NamedAttribute srcAlignAttr = rewriter.getNamedAttr( |
| 400 | name: mlir::LLVM::LLVMDialect::getAlignAttrName(), |
| 401 | val: rewriter.getI64IntegerAttr(value: op.getSrcAlignment().value_or(u&: srcTypeAlign))); |
| 402 | mlir::ArrayAttr argAttrs = rewriter.getArrayAttr(value: { |
| 403 | /*dst_attrs=*/rewriter.getDictionaryAttr(value: {dstAlignAttr}), |
| 404 | /*src_attrs=*/rewriter.getDictionaryAttr(value: {srcAlignAttr}), |
| 405 | }); |
| 406 | |
| 407 | rewriter.replaceOpWithNewOp<mlir::LLVM::MemcpyOp>( |
| 408 | op, args: adaptor.getDst(), args: adaptor.getSrc(), args: length, args: op.getIsVolatile(), |
| 409 | /*tbaa_struct=*/args: nullptr, /*access_groups=*/args: nullptr, |
| 410 | /*alias_scopes=*/args: nullptr, /*noalias_scopes=*/args: nullptr, /*tbaa=*/args: nullptr, |
| 411 | /*arg_attrs=*/args&: argAttrs, /*res_attrs=*/args: nullptr); |
| 412 | return mlir::success(); |
| 413 | } |
| 414 | |
| 415 | template <typename CIROp, typename LLVMOp> |
| 416 | static mlir::LogicalResult |
| 417 | lowerMemTransferOp(CIROp op, typename CIROp::Adaptor adaptor, |
| 418 | mlir::ConversionPatternRewriter &rewriter) { |
| 419 | mlir::ArrayAttr argAttrs; |
| 420 | if (op.getDstAlignment() || op.getSrcAlignment()) { |
| 421 | mlir::NamedAttribute dstAlignAttr = rewriter.getNamedAttr( |
| 422 | name: mlir::LLVM::LLVMDialect::getAlignAttrName(), |
| 423 | val: rewriter.getI64IntegerAttr(value: op.getDstAlignment().value_or(1))); |
| 424 | mlir::NamedAttribute srcAlignAttr = rewriter.getNamedAttr( |
| 425 | name: mlir::LLVM::LLVMDialect::getAlignAttrName(), |
| 426 | val: rewriter.getI64IntegerAttr(value: op.getSrcAlignment().value_or(1))); |
| 427 | argAttrs = rewriter.getArrayAttr(value: { |
| 428 | /*dst_attrs=*/rewriter.getDictionaryAttr(value: {dstAlignAttr}), |
| 429 | /*src_attrs=*/rewriter.getDictionaryAttr(value: {srcAlignAttr}), |
| 430 | }); |
| 431 | } |
| 432 | rewriter.replaceOpWithNewOp<LLVMOp>( |
| 433 | op, adaptor.getDst(), adaptor.getSrc(), adaptor.getLen(), |
| 434 | /*isVolatile=*/false, /*tbaa_struct=*/nullptr, |
| 435 | /*access_groups=*/nullptr, /*alias_scopes=*/nullptr, |
| 436 | /*noalias_scopes=*/nullptr, /*tbaa=*/nullptr, /*arg_attrs=*/argAttrs, |
| 437 | /*res_attrs=*/nullptr); |
| 438 | return mlir::success(); |
| 439 | } |
| 440 | |
| 441 | mlir::LogicalResult CIRToLLVMMemCpyOpLowering::matchAndRewrite( |
| 442 | cir::MemCpyOp op, OpAdaptor adaptor, |
| 443 | mlir::ConversionPatternRewriter &rewriter) const { |
| 444 | return lowerMemTransferOp<cir::MemCpyOp, mlir::LLVM::MemcpyOp>(op, adaptor, |
| 445 | rewriter); |
| 446 | } |
| 447 | |
| 448 | mlir::LogicalResult CIRToLLVMMemMoveOpLowering::matchAndRewrite( |
| 449 | cir::MemMoveOp op, OpAdaptor adaptor, |
| 450 | mlir::ConversionPatternRewriter &rewriter) const { |
| 451 | return lowerMemTransferOp<cir::MemMoveOp, mlir::LLVM::MemmoveOp>(op, adaptor, |
| 452 | rewriter); |
| 453 | } |
| 454 | |
| 455 | mlir::LogicalResult CIRToLLVMMemSetOpLowering::matchAndRewrite( |
| 456 | cir::MemSetOp op, OpAdaptor adaptor, |
| 457 | mlir::ConversionPatternRewriter &rewriter) const { |
| 458 | |
| 459 | // Capture the alignment before the replace erases `op`. |
| 460 | mlir::Attribute alignAttr = op.getAlignmentAttr(); |
| 461 | auto memset = rewriter.replaceOpWithNewOp<mlir::LLVM::MemsetOp>( |
| 462 | op, args: adaptor.getDst(), args: adaptor.getVal(), args: adaptor.getLen(), |
| 463 | /*isVolatile=*/args: false); |
| 464 | |
| 465 | if (alignAttr) { |
| 466 | // Construct a list full of empty attributes. |
| 467 | llvm::SmallVector<mlir::Attribute> attrs{memset.getNumOperands(), |
| 468 | rewriter.getDictionaryAttr(value: {})}; |
| 469 | llvm::SmallVector<mlir::NamedAttribute> destAttrs; |
| 470 | destAttrs.push_back( |
| 471 | Elt: {mlir::LLVM::LLVMDialect::getAlignAttrName(), alignAttr}); |
| 472 | attrs[memset.odsIndex_dst] = rewriter.getDictionaryAttr(value: destAttrs); |
| 473 | |
| 474 | auto arrayAttr = rewriter.getArrayAttr(value: attrs); |
| 475 | memset.setArgAttrsAttr(arrayAttr); |
| 476 | } |
| 477 | |
| 478 | return mlir::success(); |
| 479 | } |
| 480 | |
| 481 | static mlir::Value getLLVMIntCast(mlir::ConversionPatternRewriter &rewriter, |
| 482 | mlir::Value llvmSrc, mlir::Type llvmDstIntTy, |
| 483 | bool isUnsigned, uint64_t cirSrcWidth, |
| 484 | uint64_t cirDstIntWidth) { |
| 485 | if (cirSrcWidth == cirDstIntWidth) |
| 486 | return llvmSrc; |
| 487 | |
| 488 | auto loc = llvmSrc.getLoc(); |
| 489 | if (cirSrcWidth < cirDstIntWidth) { |
| 490 | if (isUnsigned) |
| 491 | return mlir::LLVM::ZExtOp::create(builder&: rewriter, location: loc, res: llvmDstIntTy, arg: llvmSrc); |
| 492 | return mlir::LLVM::SExtOp::create(builder&: rewriter, location: loc, res: llvmDstIntTy, arg: llvmSrc); |
| 493 | } |
| 494 | |
| 495 | // Otherwise truncate |
| 496 | return mlir::LLVM::TruncOp::create(builder&: rewriter, location: loc, res: llvmDstIntTy, arg: llvmSrc); |
| 497 | } |
| 498 | |
| 499 | class CIRAttrToValue { |
| 500 | public: |
| 501 | CIRAttrToValue(mlir::Location parentLoc, mlir::ModuleOp parentModule, |
| 502 | mlir::ConversionPatternRewriter &rewriter, |
| 503 | mlir::SymbolTableCollection &symbolTables, |
| 504 | const mlir::TypeConverter *converter, |
| 505 | const CIRPreConversionTypeMap &preConversionTypes, |
| 506 | LLVMBlockAddressInfo *blockInfoAddr = nullptr) |
| 507 | : parentLoc(parentLoc), parentModule(parentModule), rewriter(rewriter), |
| 508 | symbolTables(symbolTables), converter(converter), |
| 509 | preConversionTypes(preConversionTypes), blockInfoAddr(blockInfoAddr) {} |
| 510 | |
| 511 | #define GET_CIR_ATTR_TO_VALUE_VISITOR_DECLS |
| 512 | #include "clang/CIR/Dialect/IR/CIRLowering.inc" |
| 513 | #undef GET_CIR_ATTR_TO_VALUE_VISITOR_DECLS |
| 514 | |
| 515 | private: |
| 516 | struct LoweredGlobalSymbol { |
| 517 | mlir::Value addr; |
| 518 | mlir::Type sourceType; |
| 519 | mlir::Type originalType; |
| 520 | unsigned addrSpace; |
| 521 | }; |
| 522 | |
| 523 | LoweredGlobalSymbol getAddressOfSymbol(mlir::FlatSymbolRefAttr symbol); |
| 524 | mlir::Value castGlobalAddrToType(mlir::Value addr, mlir::Type attrType, |
| 525 | mlir::Type sourceType, unsigned addrSpace); |
| 526 | |
| 527 | mlir::Location parentLoc; |
| 528 | mlir::ModuleOp parentModule; |
| 529 | mlir::ConversionPatternRewriter &rewriter; |
| 530 | mlir::SymbolTableCollection &symbolTables; |
| 531 | const mlir::TypeConverter *converter; |
| 532 | const CIRPreConversionTypeMap &preConversionTypes; |
| 533 | // Only available when lowering global initializers that may contain block |
| 534 | // address attributes. Used to resolve a BlockAddrInfoAttr to its block tag. |
| 535 | LLVMBlockAddressInfo *blockInfoAddr; |
| 536 | }; |
| 537 | |
| 538 | /// Switches on the type of attribute and calls the appropriate conversion. |
| 539 | mlir::Value |
| 540 | lowerCirAttrAsValue(mlir::Location parentLoc, mlir::ModuleOp parentModule, |
| 541 | const mlir::Attribute attr, |
| 542 | mlir::ConversionPatternRewriter &rewriter, |
| 543 | mlir::SymbolTableCollection &symbolTables, |
| 544 | const mlir::TypeConverter *converter, |
| 545 | const CIRPreConversionTypeMap &preConversionTypes, |
| 546 | LLVMBlockAddressInfo *blockInfoAddr) { |
| 547 | CIRAttrToValue valueConverter(parentLoc, parentModule, rewriter, symbolTables, |
| 548 | converter, preConversionTypes, blockInfoAddr); |
| 549 | return valueConverter.visit(attr); |
| 550 | } |
| 551 | |
| 552 | static mlir::LLVM::ModRefInfo convertModRefInfo(cir::ModRefInfo info) { |
| 553 | switch (info) { |
| 554 | case cir::ModRefInfo::NoModRef: |
| 555 | return mlir::LLVM::ModRefInfo::NoModRef; |
| 556 | case cir::ModRefInfo::Ref: |
| 557 | return mlir::LLVM::ModRefInfo::Ref; |
| 558 | case cir::ModRefInfo::Mod: |
| 559 | return mlir::LLVM::ModRefInfo::Mod; |
| 560 | case cir::ModRefInfo::ModRef: |
| 561 | return mlir::LLVM::ModRefInfo::ModRef; |
| 562 | } |
| 563 | llvm_unreachable("unhandled cir::ModRefInfo" ); |
| 564 | } |
| 565 | |
| 566 | /// A null input stays null, which is how both dialects spell unknown effects. |
| 567 | static mlir::LLVM::MemoryEffectsAttr |
| 568 | convertMemoryEffects(mlir::MLIRContext *ctx, cir::MemoryEffectsAttr effects) { |
| 569 | if (!effects) |
| 570 | return {}; |
| 571 | |
| 572 | return mlir::LLVM::MemoryEffectsAttr::get( |
| 573 | context: ctx, other: convertModRefInfo(info: effects.getOther()), |
| 574 | argMem: convertModRefInfo(info: effects.getArgMem()), |
| 575 | inaccessibleMem: convertModRefInfo(info: effects.getInaccessibleMem()), |
| 576 | errnoMem: convertModRefInfo(info: effects.getErrnoMem()), |
| 577 | targetMem0: convertModRefInfo(info: effects.getTargetMem0()), |
| 578 | targetMem1: convertModRefInfo(info: effects.getTargetMem1())); |
| 579 | } |
| 580 | |
| 581 | static void convertCallEffects(mlir::Operation *callOp, bool isNothrow, |
| 582 | cir::MemoryEffectsAttr effects, |
| 583 | mlir::LLVM::MemoryEffectsAttr &memoryEffect, |
| 584 | bool &noUnwind, bool &willReturn, |
| 585 | bool &noReturn) { |
| 586 | memoryEffect = convertMemoryEffects(ctx: callOp->getContext(), effects); |
| 587 | // CIR spells two separate cannot-unwind facts, nothrow for a callee |
| 588 | // declared not to throw and nounwind for one whose effects rule out |
| 589 | // unwinding. LLVM has only nounwind, so either one sets it. |
| 590 | noUnwind = isNothrow || callOp->hasAttr(name: CIRDialect::getNoUnwindAttrName()); |
| 591 | willReturn = callOp->hasAttr(name: CIRDialect::getWillReturnAttrName()); |
| 592 | noReturn = callOp->hasAttr(name: CIRDialect::getNoReturnAttrName()); |
| 593 | } |
| 594 | |
| 595 | static mlir::LLVM::CallIntrinsicOp |
| 596 | createCallLLVMIntrinsicOp(mlir::ConversionPatternRewriter &rewriter, |
| 597 | mlir::Location loc, const llvm::Twine &intrinsicName, |
| 598 | mlir::Type resultTy, mlir::ValueRange operands, |
| 599 | mlir::LLVM::FastmathFlags fastmathFlags = {}) { |
| 600 | auto intrinsicNameAttr = |
| 601 | mlir::StringAttr::get(context: rewriter.getContext(), bytes: intrinsicName); |
| 602 | auto fmfAttr = |
| 603 | mlir::LLVM::FastmathFlagsAttr::get(context: rewriter.getContext(), value: fastmathFlags); |
| 604 | // CallIntrinsicOp has distinct void / result create overloads. The FMF |
| 605 | // builders take a TypeRange for results. |
| 606 | if (resultTy) |
| 607 | return mlir::LLVM::CallIntrinsicOp::create( |
| 608 | builder&: rewriter, location: loc, resultTypes: mlir::TypeRange{resultTy}, intrin: intrinsicNameAttr, args: operands, |
| 609 | fastMathFlags: fmfAttr); |
| 610 | return mlir::LLVM::CallIntrinsicOp::create(builder&: rewriter, location: loc, intrin: intrinsicNameAttr, |
| 611 | args: operands, fastMathFlags: fmfAttr); |
| 612 | } |
| 613 | |
| 614 | static mlir::LLVM::CallIntrinsicOp replaceOpWithCallLLVMIntrinsicOp( |
| 615 | mlir::ConversionPatternRewriter &rewriter, mlir::Operation *op, |
| 616 | const llvm::Twine &intrinsicName, mlir::Type resultTy, |
| 617 | mlir::ValueRange operands, mlir::LLVM::FastmathFlags fastmathFlags = {}) { |
| 618 | mlir::LLVM::CallIntrinsicOp callIntrinOp = createCallLLVMIntrinsicOp( |
| 619 | rewriter, loc: op->getLoc(), intrinsicName, resultTy, operands, fastmathFlags); |
| 620 | rewriter.replaceOp(op, newOp: callIntrinOp.getOperation()); |
| 621 | return callIntrinOp; |
| 622 | } |
| 623 | |
| 624 | static llvm::StringRef getConstrainedRoundingMetadata(cir::FenvAttr fenv) { |
| 625 | std::optional<cir::FPDynamicRoundingMode> rounding = |
| 626 | fenv.getDynamicRoundingMode(); |
| 627 | if (!rounding) |
| 628 | return "round.tonearest" ; |
| 629 | switch (*rounding) { |
| 630 | case cir::FPDynamicRoundingMode::ToNearest: |
| 631 | return "round.tonearest" ; |
| 632 | case cir::FPDynamicRoundingMode::Downward: |
| 633 | return "round.downward" ; |
| 634 | case cir::FPDynamicRoundingMode::Upward: |
| 635 | return "round.upward" ; |
| 636 | case cir::FPDynamicRoundingMode::UpwardZero: |
| 637 | return "round.towardzero" ; |
| 638 | case cir::FPDynamicRoundingMode::ToNearestAway: |
| 639 | return "round.tonearestaway" ; |
| 640 | case cir::FPDynamicRoundingMode::Unknown: |
| 641 | return "round.dynamic" ; |
| 642 | } |
| 643 | llvm_unreachable("unknown FP dynamic rounding mode" ); |
| 644 | } |
| 645 | |
| 646 | static llvm::StringRef getConstrainedExceptMetadata(cir::FenvAttr fenv) { |
| 647 | std::optional<cir::FPExceptionMode> exceptMode = fenv.getExceptMode(); |
| 648 | if (exceptMode == cir::FPExceptionMode::Masked) |
| 649 | return "fpexcept.ignore" ; |
| 650 | mlir::BoolAttr strictExcept = fenv.getStrictExcept(); |
| 651 | if (!strictExcept) |
| 652 | return "fpexcept.ignore" ; |
| 653 | return strictExcept.getValue() ? "fpexcept.strict" : "fpexcept.maytrap" ; |
| 654 | } |
| 655 | |
| 656 | static mlir::Value |
| 657 | createFenvMetadataValue(mlir::ConversionPatternRewriter &rewriter, |
| 658 | mlir::Location loc, llvm::StringRef str) { |
| 659 | auto mdString = mlir::LLVM::MDStringAttr::get( |
| 660 | context: rewriter.getContext(), value: mlir::StringAttr::get(context: rewriter.getContext(), bytes: str)); |
| 661 | return mlir::LLVM::MetadataAsValueOp::create(builder&: rewriter, location: loc, metadata: mdString); |
| 662 | } |
| 663 | |
| 664 | mlir::LogicalResult lowerToConstrainedFPIntrinsic( |
| 665 | mlir::Operation *op, mlir::ValueRange operands, cir::FenvAttr fenv, |
| 666 | mlir::Type llvmResTy, mlir::ConversionPatternRewriter &rewriter, |
| 667 | llvm::StringRef constrainedMnemonic, bool hasRoundingMode, |
| 668 | mlir::LLVM::FastmathFlags fastmathFlags) { |
| 669 | mlir::Location loc = op->getLoc(); |
| 670 | llvm::SmallVector<mlir::Value> callOperands(operands.begin(), operands.end()); |
| 671 | if (hasRoundingMode) |
| 672 | callOperands.push_back(Elt: createFenvMetadataValue( |
| 673 | rewriter, loc, str: getConstrainedRoundingMetadata(fenv))); |
| 674 | callOperands.push_back(Elt: createFenvMetadataValue( |
| 675 | rewriter, loc, str: getConstrainedExceptMetadata(fenv))); |
| 676 | |
| 677 | replaceOpWithCallLLVMIntrinsicOp( |
| 678 | rewriter, op, intrinsicName: "llvm.experimental.constrained." + constrainedMnemonic, |
| 679 | resultTy: llvmResTy, operands: callOperands, fastmathFlags); |
| 680 | return mlir::success(); |
| 681 | } |
| 682 | |
| 683 | static mlir::LLVM::FastmathFlags |
| 684 | convertFastMathFlags(cir::FastMathFlags cirFlags); |
| 685 | |
| 686 | static mlir::Value createConstrainedFPIntrinsicCall( |
| 687 | mlir::ConversionPatternRewriter &rewriter, mlir::Location loc, |
| 688 | mlir::ValueRange operands, mlir::Type llvmResTy, cir::FenvAttr fenv, |
| 689 | llvm::StringRef constrainedMnemonic, bool hasRoundingMode, |
| 690 | mlir::LLVM::FastmathFlags fastmathFlags = {}) { |
| 691 | llvm::SmallVector<mlir::Value> callOperands(operands.begin(), operands.end()); |
| 692 | if (hasRoundingMode) |
| 693 | callOperands.push_back(Elt: createFenvMetadataValue( |
| 694 | rewriter, loc, str: getConstrainedRoundingMetadata(fenv))); |
| 695 | |
| 696 | callOperands.push_back(Elt: createFenvMetadataValue( |
| 697 | rewriter, loc, str: getConstrainedExceptMetadata(fenv))); |
| 698 | |
| 699 | mlir::LLVM::CallIntrinsicOp intrinsic = createCallLLVMIntrinsicOp( |
| 700 | rewriter, loc, intrinsicName: "llvm.experimental.constrained." + constrainedMnemonic, |
| 701 | resultTy: llvmResTy, operands: callOperands, fastmathFlags); |
| 702 | return intrinsic->getResult(idx: 0); |
| 703 | } |
| 704 | |
| 705 | template <typename LLVMOp> |
| 706 | mlir::LogicalResult lowerFPOp(mlir::Operation *op, mlir::ValueRange operands, |
| 707 | cir::FastMathFlagsAttr fastmath, |
| 708 | const mlir::TypeConverter &typeConverter, |
| 709 | mlir::ConversionPatternRewriter &rewriter) { |
| 710 | mlir::Type llvmResTy = typeConverter.convertType(t: op->getResultTypes()[0]); |
| 711 | if (!llvmResTy) |
| 712 | return op->emitError(message: "expected LLVM result type for floating-point op" ); |
| 713 | |
| 714 | auto newOp = rewriter.replaceOpWithNewOp<LLVMOp>( |
| 715 | op, mlir::TypeRange{llvmResTy}, operands, |
| 716 | cir::getDefaultProperties<LLVMOp>(op->getContext())); |
| 717 | if (fastmath) |
| 718 | mlir::cast<mlir::LLVM::FastmathFlagsInterface>(newOp.getOperation()) |
| 719 | .setFastmathAttr(mlir::LLVM::FastmathFlagsAttr::get( |
| 720 | context: rewriter.getContext(), value: convertFastMathFlags(cirFlags: fastmath.getValue()))); |
| 721 | return mlir::success(); |
| 722 | } |
| 723 | |
| 724 | template <typename LLVMOp> |
| 725 | mlir::LogicalResult lowerConstrainableFPOp( |
| 726 | mlir::Operation *op, mlir::ValueRange operands, cir::FenvAttr fenv, |
| 727 | cir::FastMathFlagsAttr fastmath, const mlir::TypeConverter &typeConverter, |
| 728 | mlir::ConversionPatternRewriter &rewriter, |
| 729 | llvm::StringRef constrainedMnemonic, bool hasRoundingMode) { |
| 730 | if (!fenv) |
| 731 | return lowerFPOp<LLVMOp>(op, operands, fastmath, typeConverter, rewriter); |
| 732 | |
| 733 | mlir::Type llvmResTy = typeConverter.convertType(t: op->getResultTypes()[0]); |
| 734 | if (!llvmResTy) |
| 735 | return op->emitError(message: "expected LLVM result type for floating-point op" ); |
| 736 | |
| 737 | mlir::LLVM::FastmathFlags fastmathFlags = {}; |
| 738 | if (fastmath) |
| 739 | fastmathFlags = convertFastMathFlags(cirFlags: fastmath.getValue()); |
| 740 | |
| 741 | return lowerToConstrainedFPIntrinsic(op, operands, fenv, llvmResTy, rewriter, |
| 742 | constrainedMnemonic, hasRoundingMode, |
| 743 | fastmathFlags); |
| 744 | } |
| 745 | |
| 746 | mlir::LLVM::FastmathFlagsAttr |
| 747 | convertToLLVMProperty(mlir::MLIRContext *ctx, |
| 748 | std::optional<cir::FastMathFlags> flags) { |
| 749 | return mlir::LLVM::FastmathFlagsAttr::get( |
| 750 | context: ctx, value: flags ? convertFastMathFlags(cirFlags: *flags) : mlir::LLVM::FastmathFlags{}); |
| 751 | } |
| 752 | |
| 753 | mlir::IntegerAttr convertToLLVMProperty(mlir::MLIRContext *ctx, |
| 754 | cir::FPClassTest flags) { |
| 755 | return mlir::IntegerAttr::get(type: mlir::IntegerType::get(context: ctx, width: 32), |
| 756 | value: static_cast<uint32_t>(flags)); |
| 757 | } |
| 758 | |
| 759 | mlir::IntegerAttr convertToLLVMProperty(mlir::MLIRContext *ctx, bool flag) { |
| 760 | return mlir::IntegerAttr::get(type: mlir::IntegerType::get(context: ctx, width: 1), value: flag); |
| 761 | } |
| 762 | |
| 763 | static mlir::LLVM::FastmathFlags |
| 764 | convertFastMathFlags(cir::FastMathFlags cirFlags) { |
| 765 | mlir::LLVM::FastmathFlags llvmFlags{}; |
| 766 | const std::pair<cir::FastMathFlags, mlir::LLVM::FastmathFlags> flags[] = { |
| 767 | {cir::FastMathFlags::nnan, mlir::LLVM::FastmathFlags::nnan}, |
| 768 | {cir::FastMathFlags::ninf, mlir::LLVM::FastmathFlags::ninf}, |
| 769 | {cir::FastMathFlags::nsz, mlir::LLVM::FastmathFlags::nsz}, |
| 770 | {cir::FastMathFlags::arcp, mlir::LLVM::FastmathFlags::arcp}, |
| 771 | {cir::FastMathFlags::contract, mlir::LLVM::FastmathFlags::contract}, |
| 772 | {cir::FastMathFlags::afn, mlir::LLVM::FastmathFlags::afn}, |
| 773 | {cir::FastMathFlags::reassoc, mlir::LLVM::FastmathFlags::reassoc}, |
| 774 | }; |
| 775 | |
| 776 | for (auto [cirFlag, llvmFlag] : flags) { |
| 777 | if (bitEnumContainsAny(bits: cirFlags, bit: cirFlag)) |
| 778 | llvmFlags = llvmFlags | llvmFlag; |
| 779 | } |
| 780 | |
| 781 | return llvmFlags; |
| 782 | } |
| 783 | |
| 784 | mlir::LogicalResult CIRToLLVMLLVMIntrinsicCallOpLowering::matchAndRewrite( |
| 785 | cir::LLVMIntrinsicCallOp op, OpAdaptor adaptor, |
| 786 | mlir::ConversionPatternRewriter &rewriter) const { |
| 787 | // Result is Optional on the op, so void intrinsics have zero |
| 788 | // results; leave llvmResTy null in that case. |
| 789 | mlir::Type llvmResTy; |
| 790 | if (op->getNumResults() != 0) { |
| 791 | llvmResTy = getTypeConverter()->convertType(t: op->getResultTypes()[0]); |
| 792 | if (!llvmResTy) |
| 793 | return op.emitError(message: "expected LLVM result type" ); |
| 794 | } |
| 795 | StringRef name = op.getIntrinsicName(); |
| 796 | mlir::LLVM::FastmathFlags fastmathFlags = {}; |
| 797 | if (std::optional<cir::FastMathFlags> fastmath = op.getFastmathFlags()) |
| 798 | fastmathFlags = convertFastMathFlags(cirFlags: *fastmath); |
| 799 | |
| 800 | // Some LLVM intrinsics require ElementType attribute to be attached to |
| 801 | // the argument of pointer type. That prevents us from generating LLVM IR |
| 802 | // because from LLVM dialect, we have LLVM IR like the below which fails |
| 803 | // LLVM IR verification. |
| 804 | // %3 = call i64 @llvm.aarch64.ldxr.p0(ptr %2) |
| 805 | // The expected LLVM IR should be like |
| 806 | // %3 = call i64 @llvm.aarch64.ldxr.p0(ptr elementtype(i32) %2) |
| 807 | // TODO(cir): MLIR LLVM dialect should handle this part as CIR has no way |
| 808 | // to set LLVM IR attribute. |
| 809 | assert(!cir::MissingFeatures::intrinsicElementTypeSupport()); |
| 810 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm." + name, resultTy: llvmResTy, |
| 811 | operands: adaptor.getOperands(), fastmathFlags); |
| 812 | return mlir::success(); |
| 813 | } |
| 814 | |
| 815 | /// BoolAttr visitor. |
| 816 | mlir::Value CIRAttrToValue::visitCirAttr(cir::BoolAttr boolAttr) { |
| 817 | mlir::DataLayout layout(parentModule); |
| 818 | mlir::Value boolVal = mlir::LLVM::ConstantOp::create( |
| 819 | builder&: rewriter, location: parentLoc, type: converter->convertType(t: boolAttr.getType()), |
| 820 | value: boolAttr.getValue()); |
| 821 | return emitToMemory(rewriter, dataLayout: layout, origType: boolAttr.getType(), value: boolVal); |
| 822 | } |
| 823 | |
| 824 | /// IntAttr visitor. |
| 825 | mlir::Value CIRAttrToValue::visitCirAttr(cir::IntAttr intAttr) { |
| 826 | mlir::DataLayout layout(parentModule); |
| 827 | |
| 828 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val: intAttr.getType()); |
| 829 | intTy && intTy.isBitInt()) { |
| 830 | mlir::Type biTy = convertTypeForMemory(converter: *converter, dataLayout: layout, type: intTy); |
| 831 | return mlir::LLVM::ConstantOp::create( |
| 832 | builder&: rewriter, location: parentLoc, res: biTy, |
| 833 | value: getBitIntStorageAttr(rewriter, attr: intAttr, dataLayout: layout)); |
| 834 | } |
| 835 | |
| 836 | // Materialize the value at its literal width, then widen to the in-memory |
| 837 | // storage type (a no-op except for _BitInt) so aggregate members built here |
| 838 | // match the iM struct/array fields produced by convertTypeForMemory. |
| 839 | mlir::Value val = mlir::LLVM::ConstantOp::create( |
| 840 | builder&: rewriter, location: parentLoc, type: converter->convertType(t: intAttr.getType()), |
| 841 | value: intAttr.getValue()); |
| 842 | return emitToMemory(rewriter, dataLayout: layout, origType: intAttr.getType(), value: val); |
| 843 | } |
| 844 | |
| 845 | /// FPAttr visitor. |
| 846 | mlir::Value CIRAttrToValue::visitCirAttr(cir::FPAttr fltAttr) { |
| 847 | return mlir::LLVM::ConstantOp::create( |
| 848 | builder&: rewriter, location: parentLoc, type: converter->convertType(t: fltAttr.getType()), |
| 849 | value: fltAttr.getValue()); |
| 850 | } |
| 851 | |
| 852 | /// ConstComplexAttr visitor. |
| 853 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ConstComplexAttr complexAttr) { |
| 854 | cir::ComplexType complexType = complexAttr.getType(); |
| 855 | mlir::Type complexElemTy = complexType.getElementType(); |
| 856 | mlir::Type complexElemLLVMTy = converter->convertType(t: complexElemTy); |
| 857 | |
| 858 | mlir::Attribute components[2]; |
| 859 | if (const auto intType = mlir::dyn_cast<cir::IntType>(Val&: complexElemTy)) { |
| 860 | components[0] = rewriter.getIntegerAttr( |
| 861 | type: complexElemLLVMTy, |
| 862 | value: mlir::cast<cir::IntAttr>(Val: complexAttr.getReal()).getValue()); |
| 863 | components[1] = rewriter.getIntegerAttr( |
| 864 | type: complexElemLLVMTy, |
| 865 | value: mlir::cast<cir::IntAttr>(Val: complexAttr.getImag()).getValue()); |
| 866 | } else { |
| 867 | components[0] = rewriter.getFloatAttr( |
| 868 | type: complexElemLLVMTy, |
| 869 | value: mlir::cast<cir::FPAttr>(Val: complexAttr.getReal()).getValue()); |
| 870 | components[1] = rewriter.getFloatAttr( |
| 871 | type: complexElemLLVMTy, |
| 872 | value: mlir::cast<cir::FPAttr>(Val: complexAttr.getImag()).getValue()); |
| 873 | } |
| 874 | |
| 875 | return mlir::LLVM::ConstantOp::create( |
| 876 | builder&: rewriter, location: parentLoc, res: converter->convertType(t: complexAttr.getType()), |
| 877 | value: rewriter.getArrayAttr(value: components)); |
| 878 | } |
| 879 | |
| 880 | /// ConstPtrAttr visitor. |
| 881 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ConstPtrAttr ptrAttr) { |
| 882 | if (ptrAttr.isNullValue()) { |
| 883 | return mlir::LLVM::ZeroOp::create( |
| 884 | builder&: rewriter, location: parentLoc, res: converter->convertType(t: ptrAttr.getType())); |
| 885 | } |
| 886 | mlir::DataLayout layout(parentModule); |
| 887 | mlir::Value ptrVal = mlir::LLVM::ConstantOp::create( |
| 888 | builder&: rewriter, location: parentLoc, |
| 889 | type: rewriter.getIntegerType(width: layout.getTypeSizeInBits(t: ptrAttr.getType())), |
| 890 | value: ptrAttr.getValue().getInt()); |
| 891 | return mlir::LLVM::IntToPtrOp::create( |
| 892 | builder&: rewriter, location: parentLoc, resultType: converter->convertType(t: ptrAttr.getType()), operands: ptrVal); |
| 893 | } |
| 894 | |
| 895 | /// BlockAddrInfoAttr visitor. |
| 896 | mlir::Value CIRAttrToValue::visitCirAttr(cir::BlockAddrInfoAttr blockAddrInfo) { |
| 897 | assert(blockInfoAddr && |
| 898 | "block address lowering requires LLVMBlockAddressInfo" ); |
| 899 | // A block address is lowered to an llvm.blockaddress op that references a |
| 900 | // block tag inside the target function. The matching block tag may not have |
| 901 | // been emitted yet, in which case the address is recorded as unresolved and |
| 902 | // patched up later in resolveBlockAddressOp. |
| 903 | mlir::LLVM::BlockTagOp matchLabel = |
| 904 | blockInfoAddr->lookupBlockTag(info: blockAddrInfo); |
| 905 | mlir::LLVM::BlockTagAttr tagAttr = |
| 906 | matchLabel ? matchLabel.getTag() : mlir::LLVM::BlockTagAttr{}; |
| 907 | auto blkAddr = mlir::LLVM::BlockAddressAttr::get( |
| 908 | context: rewriter.getContext(), function: blockAddrInfo.getFunc(), tag: tagAttr); |
| 909 | auto blockAddressOp = mlir::LLVM::BlockAddressOp::create( |
| 910 | builder&: rewriter, location: parentLoc, |
| 911 | res: mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()), block_addr: blkAddr); |
| 912 | if (!matchLabel) |
| 913 | blockInfoAddr->addUnresolvedBlockAddress(op: blockAddressOp, info: blockAddrInfo); |
| 914 | return blockAddressOp; |
| 915 | } |
| 916 | |
| 917 | /// BlockAddrDiffAttr visitor. |
| 918 | mlir::Value CIRAttrToValue::visitCirAttr(cir::BlockAddrDiffAttr blockAddrDiff) { |
| 919 | assert(blockInfoAddr && |
| 920 | "block address lowering requires LLVMBlockAddressInfo" ); |
| 921 | // A block-address difference initializer is lowered to the difference of the |
| 922 | // two block addresses: trunc(ptrtoint(lhs) - ptrtoint(rhs)). Just like a |
| 923 | // single block address, each referenced block tag may not have been emitted |
| 924 | // yet, in which case it is recorded as unresolved and patched up later in |
| 925 | // resolveBlockAddressOp. |
| 926 | mlir::DataLayout layout(parentModule); |
| 927 | mlir::MLIRContext *ctx = rewriter.getContext(); |
| 928 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: ctx); |
| 929 | |
| 930 | auto emitBlockAddr = [&](mlir::StringAttr label) -> mlir::Value { |
| 931 | auto info = cir::BlockAddrInfoAttr::get( |
| 932 | context: ctx, func_name: blockAddrDiff.getFunc().getValue(), label_name: label.getValue()); |
| 933 | mlir::LLVM::BlockTagOp matchLabel = blockInfoAddr->lookupBlockTag(info); |
| 934 | mlir::LLVM::BlockTagAttr tagAttr = |
| 935 | matchLabel ? matchLabel.getTag() : mlir::LLVM::BlockTagAttr{}; |
| 936 | auto blkAddr = mlir::LLVM::BlockAddressAttr::get( |
| 937 | context: ctx, function: blockAddrDiff.getFunc(), tag: tagAttr); |
| 938 | auto addrOp = |
| 939 | mlir::LLVM::BlockAddressOp::create(builder&: rewriter, location: parentLoc, res: ptrTy, block_addr: blkAddr); |
| 940 | if (!matchLabel) |
| 941 | blockInfoAddr->addUnresolvedBlockAddress(op: addrOp, info); |
| 942 | return addrOp; |
| 943 | }; |
| 944 | |
| 945 | mlir::Value lhsAddr = emitBlockAddr(blockAddrDiff.getLhsLabel()); |
| 946 | mlir::Value rhsAddr = emitBlockAddr(blockAddrDiff.getRhsLabel()); |
| 947 | |
| 948 | // Compute the difference in a pointer-sized integer, then truncate to the |
| 949 | // initializer's type. LLVM is sensitive about the exact format of the |
| 950 | // address-of-label difference, so the truncation must happen after the |
| 951 | // subtraction. |
| 952 | mlir::Type intptrTy = |
| 953 | rewriter.getIntegerType(width: layout.getTypeSizeInBits(t: ptrTy)); |
| 954 | mlir::Value lhsInt = |
| 955 | mlir::LLVM::PtrToIntOp::create(builder&: rewriter, location: parentLoc, res: intptrTy, arg: lhsAddr); |
| 956 | mlir::Value rhsInt = |
| 957 | mlir::LLVM::PtrToIntOp::create(builder&: rewriter, location: parentLoc, res: intptrTy, arg: rhsAddr); |
| 958 | mlir::Value diffVal = |
| 959 | mlir::LLVM::SubOp::create(builder&: rewriter, location: parentLoc, lhs: lhsInt, rhs: rhsInt); |
| 960 | |
| 961 | mlir::Type resultTy = converter->convertType(t: blockAddrDiff.getType()); |
| 962 | mlir::Value result = diffVal; |
| 963 | if (resultTy != intptrTy) |
| 964 | result = |
| 965 | mlir::LLVM::TruncOp::create(builder&: rewriter, location: parentLoc, res: resultTy, arg: diffVal); |
| 966 | return result; |
| 967 | } |
| 968 | |
| 969 | // ConstArrayAttr visitor |
| 970 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ConstArrayAttr attr) { |
| 971 | mlir::Type convertedTy = converter->convertType(t: attr.getType()); |
| 972 | mlir::DataLayout dataLayout(parentModule); |
| 973 | mlir::Type failedType = attr.getType(); |
| 974 | mlir::Type llvmTy = adjustGlobalTypeForInit(llvmType: convertedTy, init: attr, converter: *converter, |
| 975 | dataLayout, failedType: &failedType); |
| 976 | if (!llvmTy) { |
| 977 | emitTypeConversionError(loc: parentLoc, type: failedType); |
| 978 | return {}; |
| 979 | } |
| 980 | mlir::Value result; |
| 981 | |
| 982 | // When the array can be represented as a single constant attribute, emit one |
| 983 | // llvm.mlir.constant instead of a chain of llvm.insertvalue ops. That works |
| 984 | // for any array type, including an array of split rows, but not for a split |
| 985 | // array itself: a struct-typed constant needs scalar fields. |
| 986 | if (mlir::isa<mlir::LLVM::LLVMArrayType>(Val: llvmTy)) |
| 987 | if (std::optional<mlir::Attribute> constAttr = lowerConstArrayAttr( |
| 988 | constArr: attr, symbolTables, converter, preConversionTypes, moduleOp: parentModule)) |
| 989 | return mlir::LLVM::ConstantOp::create(builder&: rewriter, location: parentLoc, res: llvmTy, |
| 990 | value: *constAttr); |
| 991 | |
| 992 | if (attr.hasTrailingZeros()) |
| 993 | result = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: parentLoc, res: llvmTy); |
| 994 | else |
| 995 | result = mlir::LLVM::UndefOp::create(builder&: rewriter, location: parentLoc, res: llvmTy); |
| 996 | |
| 997 | // Iteratively lower each constant element of the array. |
| 998 | if (auto arrayAttr = mlir::dyn_cast<mlir::ArrayAttr>(Val: attr.getElts())) { |
| 999 | // Fewer struct members than elements: adjustGlobalTypeForInit grouped the |
| 1000 | // elements of an array ending in many zeros into one leading array. That |
| 1001 | // array has no trailing zeros, so it keeps its type and still lowers as |
| 1002 | // one dense constant. |
| 1003 | // |
| 1004 | // FIXME: A split array itself is a struct, which llvm.mlir.constant cannot |
| 1005 | // hold with aggregate fields, so on this path each split array costs |
| 1006 | // insertvalue ops. With many split arrays that cannot be folded into an |
| 1007 | // enclosing constant, that may be slower than lowering them unsplit as |
| 1008 | // dense arrays; consider a heuristic choosing between the two. |
| 1009 | auto structTy = mlir::dyn_cast<mlir::LLVM::LLVMStructType>(Val&: llvmTy); |
| 1010 | if (structTy && structTy.getBody().size() < arrayAttr.size()) { |
| 1011 | auto arrayTy = mlir::cast<cir::ArrayType>(Val: attr.getType()); |
| 1012 | auto prefix = cir::ConstArrayAttr::get( |
| 1013 | type: cir::ArrayType::get(elementType: arrayTy.getElementType(), size: arrayAttr.size()), |
| 1014 | elts: arrayAttr); |
| 1015 | mlir::Value init = visit(attr: prefix); |
| 1016 | if (!init) |
| 1017 | return {}; |
| 1018 | return mlir::LLVM::InsertValueOp::create( |
| 1019 | builder&: rewriter, location: parentLoc, container: result, value: init, position: llvm::ArrayRef<int64_t>{0}); |
| 1020 | } |
| 1021 | for (auto [idx, elt] : llvm::enumerate(First&: arrayAttr)) { |
| 1022 | mlir::Value init = visit(attr: elt); |
| 1023 | if (!init) |
| 1024 | return {}; |
| 1025 | result = mlir::LLVM::InsertValueOp::create(builder&: rewriter, location: parentLoc, container: result, |
| 1026 | value: init, position: idx); |
| 1027 | } |
| 1028 | } else if (auto strAttr = mlir::dyn_cast<mlir::StringAttr>(Val: attr.getElts())) { |
| 1029 | // TODO(cir): this diverges from traditional lowering. Normally the string |
| 1030 | // would be a global constant that is memcopied. |
| 1031 | auto arrayTy = mlir::dyn_cast<cir::ArrayType>(Val: strAttr.getType()); |
| 1032 | assert(arrayTy && "String attribute must have an array type" ); |
| 1033 | mlir::Type eltTy = arrayTy.getElementType(); |
| 1034 | for (auto [idx, elt] : llvm::enumerate(First&: strAttr)) { |
| 1035 | auto init = mlir::LLVM::ConstantOp::create( |
| 1036 | builder&: rewriter, location: parentLoc, type: converter->convertType(t: eltTy), value: elt); |
| 1037 | result = mlir::LLVM::InsertValueOp::create(builder&: rewriter, location: parentLoc, container: result, |
| 1038 | value: init, position: idx); |
| 1039 | } |
| 1040 | } else { |
| 1041 | llvm_unreachable("unexpected ConstArrayAttr elements" ); |
| 1042 | } |
| 1043 | |
| 1044 | return result; |
| 1045 | } |
| 1046 | |
| 1047 | /// ConstRecord visitor. |
| 1048 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ConstRecordAttr constRecord) { |
| 1049 | mlir::Type llvmTy = converter->convertType(t: constRecord.getType()); |
| 1050 | mlir::DataLayout dataLayout(parentModule); |
| 1051 | llvm::SmallVector<unsigned> paddingAddedIndexes; |
| 1052 | mlir::Type failedType = constRecord.getType(); |
| 1053 | llvmTy = adjustGlobalTypeForInit(llvmType: llvmTy, init: constRecord, converter: *converter, dataLayout, |
| 1054 | paddingAddedIndexes, failedType: &failedType); |
| 1055 | if (!llvmTy) { |
| 1056 | emitTypeConversionError(loc: parentLoc, type: failedType); |
| 1057 | return {}; |
| 1058 | } |
| 1059 | mlir::Value result = mlir::LLVM::UndefOp::create(builder&: rewriter, location: parentLoc, res: llvmTy); |
| 1060 | |
| 1061 | uint64_t insertIdx = 0; |
| 1062 | auto paddingItr = paddingAddedIndexes.begin(); |
| 1063 | |
| 1064 | // Iteratively lower each constant element of the record. A record member |
| 1065 | // that owns no bytes has no LLVM field and takes no element, so the elements |
| 1066 | // run one per field. |
| 1067 | for (auto [idx, elt] : llvm::enumerate(First: constRecord.getMembers())) { |
| 1068 | if (paddingItr != paddingAddedIndexes.end() && *paddingItr == idx) { |
| 1069 | ++insertIdx; |
| 1070 | ++paddingItr; |
| 1071 | } |
| 1072 | |
| 1073 | mlir::Value init = visit(attr: elt); |
| 1074 | if (!init) |
| 1075 | return {}; |
| 1076 | result = mlir::LLVM::InsertValueOp::create(builder&: rewriter, location: parentLoc, container: result, |
| 1077 | value: init, position: insertIdx); |
| 1078 | ++insertIdx; |
| 1079 | } |
| 1080 | |
| 1081 | return result; |
| 1082 | } |
| 1083 | |
| 1084 | /// ConstVectorAttr visitor. |
| 1085 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ConstVectorAttr attr) { |
| 1086 | const mlir::Type llvmTy = converter->convertType(t: attr.getType()); |
| 1087 | |
| 1088 | SmallVector<mlir::Attribute> mlirValues; |
| 1089 | for (const mlir::Attribute elementAttr : attr.getElts()) { |
| 1090 | mlir::Attribute mlirAttr; |
| 1091 | if (auto intAttr = mlir::dyn_cast<cir::IntAttr>(Val: elementAttr)) { |
| 1092 | mlirAttr = rewriter.getIntegerAttr( |
| 1093 | type: converter->convertType(t: intAttr.getType()), value: intAttr.getValue()); |
| 1094 | } else if (auto floatAttr = mlir::dyn_cast<cir::FPAttr>(Val: elementAttr)) { |
| 1095 | mlirAttr = rewriter.getFloatAttr( |
| 1096 | type: converter->convertType(t: floatAttr.getType()), value: floatAttr.getValue()); |
| 1097 | } else if (auto boolAttr = mlir::dyn_cast<cir::BoolAttr>(Val: elementAttr)) { |
| 1098 | mlirAttr = rewriter.getBoolAttr(value: boolAttr.getValue()); |
| 1099 | } else { |
| 1100 | llvm_unreachable("vector constant with an element that is neither an " |
| 1101 | "int, a float, or a bool" ); |
| 1102 | } |
| 1103 | mlirValues.push_back(Elt: mlirAttr); |
| 1104 | } |
| 1105 | |
| 1106 | return mlir::LLVM::ConstantOp::create( |
| 1107 | builder&: rewriter, location: parentLoc, res: llvmTy, |
| 1108 | value: mlir::DenseElementsAttr::get(type: mlir::cast<mlir::ShapedType>(Val: llvmTy), |
| 1109 | values: mlirValues)); |
| 1110 | } |
| 1111 | |
| 1112 | // GlobalViewAttr / GlobalOffsetAttr visitors. |
| 1113 | CIRAttrToValue::LoweredGlobalSymbol |
| 1114 | CIRAttrToValue::getAddressOfSymbol(mlir::FlatSymbolRefAttr symbol) { |
| 1115 | cir::CIRDataLayout cirDataLayout(parentModule); |
| 1116 | const mlir::DataLayout &dataLayout = cirDataLayout.layout; |
| 1117 | mlir::Type sourceType; |
| 1118 | mlir::Type originalType; |
| 1119 | unsigned addrSpace = 0; |
| 1120 | llvm::StringRef symName; |
| 1121 | mlir::Operation *sourceSymbol = |
| 1122 | symbolTables.lookupSymbolIn(symbolTableOp: parentModule, name: symbol); |
| 1123 | if (auto llvmSymbol = dyn_cast<mlir::LLVM::GlobalOp>(Val: sourceSymbol)) { |
| 1124 | if (mlir::Type preType = |
| 1125 | preConversionTypes.lookup(Key: llvmSymbol.getSymName())) { |
| 1126 | originalType = preType; |
| 1127 | sourceType = convertTypeForMemory(converter: *converter, dataLayout, type: originalType); |
| 1128 | } else { |
| 1129 | sourceType = llvmSymbol.getType(); |
| 1130 | } |
| 1131 | symName = llvmSymbol.getSymName(); |
| 1132 | addrSpace = llvmSymbol.getAddrSpace(); |
| 1133 | } else if (auto cirSymbol = dyn_cast<cir::GlobalOp>(Val: sourceSymbol)) { |
| 1134 | originalType = cirSymbol.getSymType(); |
| 1135 | sourceType = convertTypeForMemory(converter: *converter, dataLayout, type: originalType); |
| 1136 | symName = cirSymbol.getSymName(); |
| 1137 | if (auto targetAS = mlir::dyn_cast_if_present<cir::TargetAddressSpaceAttr>( |
| 1138 | Val: cirSymbol.getAddrSpaceAttr())) |
| 1139 | addrSpace = targetAS.getValue(); |
| 1140 | } else if (auto llvmFun = dyn_cast<mlir::LLVM::LLVMFuncOp>(Val: sourceSymbol)) { |
| 1141 | if (mlir::Type preType = preConversionTypes.lookup(Key: llvmFun.getSymName())) { |
| 1142 | originalType = preType; |
| 1143 | sourceType = converter->convertType(t: originalType); |
| 1144 | } else { |
| 1145 | sourceType = llvmFun.getFunctionType(); |
| 1146 | } |
| 1147 | symName = llvmFun.getSymName(); |
| 1148 | addrSpace = cirDataLayout.getProgramAddressSpace(); |
| 1149 | } else if (auto fun = dyn_cast<cir::FuncOp>(Val: sourceSymbol)) { |
| 1150 | originalType = fun.getFunctionType(); |
| 1151 | sourceType = converter->convertType(t: originalType); |
| 1152 | symName = fun.getSymName(); |
| 1153 | addrSpace = cirDataLayout.getProgramAddressSpace(); |
| 1154 | } else if (auto alias = dyn_cast<mlir::LLVM::AliasOp>(Val: sourceSymbol)) { |
| 1155 | if (mlir::Type preType = preConversionTypes.lookup(Key: alias.getSymName())) { |
| 1156 | originalType = preType; |
| 1157 | if (mlir::isa<cir::FuncType>(Val: preType)) |
| 1158 | sourceType = converter->convertType(t: originalType); |
| 1159 | else |
| 1160 | sourceType = convertTypeForMemory(converter: *converter, dataLayout, type: originalType); |
| 1161 | } else { |
| 1162 | sourceType = alias.getType(); |
| 1163 | } |
| 1164 | symName = alias.getSymName(); |
| 1165 | addrSpace = alias.getAddrSpace(); |
| 1166 | } else { |
| 1167 | llvm_unreachable("Unexpected GlobalOp type" ); |
| 1168 | } |
| 1169 | |
| 1170 | mlir::Value addr = mlir::LLVM::AddressOfOp::create( |
| 1171 | builder&: rewriter, location: parentLoc, |
| 1172 | res: mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: addrSpace), |
| 1173 | global_name: symName); |
| 1174 | return {.addr: addr, .sourceType: sourceType, .originalType: originalType, .addrSpace: addrSpace}; |
| 1175 | } |
| 1176 | |
| 1177 | mlir::Value CIRAttrToValue::castGlobalAddrToType(mlir::Value addr, |
| 1178 | mlir::Type attrType, |
| 1179 | mlir::Type sourceType, |
| 1180 | unsigned addrSpace) { |
| 1181 | |
| 1182 | // We can have a global view with an integer type in the case of method |
| 1183 | // pointers. With the Itanium ABI, the #cir.method attribute is lowered to a |
| 1184 | // #cir.global_view with a pointer-sized integer representing the address of |
| 1185 | // the method. |
| 1186 | if (mlir::isa<cir::IntType>(Val: attrType)) { |
| 1187 | mlir::Type llvmDstTy = converter->convertType(t: attrType); |
| 1188 | return mlir::LLVM::PtrToIntOp::create(builder&: rewriter, location: parentLoc, res: llvmDstTy, arg: addr); |
| 1189 | } |
| 1190 | |
| 1191 | if (auto ptrTy = mlir::dyn_cast<cir::PointerType>(Val&: attrType)) { |
| 1192 | auto llvmDstTy = converter->convertType<mlir::LLVM::LLVMPointerType>(t: ptrTy); |
| 1193 | |
| 1194 | if (llvmDstTy.getAddressSpace() != addrSpace) |
| 1195 | addr = mlir::LLVM::AddrSpaceCastOp::create(builder&: rewriter, location: parentLoc, res: llvmDstTy, |
| 1196 | arg: addr); |
| 1197 | |
| 1198 | mlir::DataLayout dataLayout(parentModule); |
| 1199 | mlir::Type llvmEltTy = |
| 1200 | convertTypeForMemory(converter: *converter, dataLayout, type: ptrTy.getPointee()); |
| 1201 | |
| 1202 | // No further cast needed if the pointee type already matches. |
| 1203 | if (llvmEltTy == sourceType) |
| 1204 | return addr; |
| 1205 | |
| 1206 | // With opaque pointers, the pointer type is already correct (either from |
| 1207 | // the original AddressOfOp or after an addrspacecast) — skip the |
| 1208 | // redundant bitcast. |
| 1209 | if (addr.getType() == llvmDstTy) |
| 1210 | return addr; |
| 1211 | |
| 1212 | return mlir::LLVM::BitcastOp::create(builder&: rewriter, location: parentLoc, res: llvmDstTy, arg: addr); |
| 1213 | } |
| 1214 | |
| 1215 | if (mlir::isa<cir::VPtrType>(Val: attrType)) |
| 1216 | return addr; |
| 1217 | |
| 1218 | llvm_unreachable("Expecting pointer or integer type" ); |
| 1219 | } |
| 1220 | |
| 1221 | mlir::Value CIRAttrToValue::visitCirAttr(cir::GlobalViewAttr globalAttr) { |
| 1222 | LoweredGlobalSymbol lowered = getAddressOfSymbol(symbol: globalAttr.getSymbol()); |
| 1223 | mlir::Value addrOp = lowered.addr; |
| 1224 | mlir::Type sourceType = lowered.sourceType; |
| 1225 | |
| 1226 | if (globalAttr.getIndices()) { |
| 1227 | if (!sourceType) { |
| 1228 | emitTypeConversionError(loc: parentLoc, type: lowered.originalType) |
| 1229 | << " for symbol " << globalAttr.getSymbol(); |
| 1230 | return {}; |
| 1231 | } |
| 1232 | llvm::SmallVector<mlir::LLVM::GEPArg> indices; |
| 1233 | |
| 1234 | if (mlir::isa<mlir::LLVM::LLVMArrayType, mlir::LLVM::LLVMStructType>( |
| 1235 | Val: sourceType)) |
| 1236 | indices.push_back(Elt: 0); |
| 1237 | |
| 1238 | for (mlir::Attribute idx : globalAttr.getIndices()) { |
| 1239 | auto intAttr = mlir::cast<mlir::IntegerAttr>(Val&: idx); |
| 1240 | indices.push_back(Elt: intAttr.getValue().getSExtValue()); |
| 1241 | } |
| 1242 | mlir::Type resTy = addrOp.getType(); |
| 1243 | mlir::Type eltTy = converter->convertType(t: sourceType); |
| 1244 | auto gep = |
| 1245 | mlir::LLVM::GEPOp::create(builder&: rewriter, location: parentLoc, resultType: resTy, elementType: eltTy, basePtr: addrOp, |
| 1246 | indices, noWrapFlags: mlir::LLVM::GEPNoWrapFlags::none); |
| 1247 | if (globalAttr.getAddressPoint()) { |
| 1248 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 1249 | if (usesRelativeVTables(mod: parentModule)) { |
| 1250 | mlir::emitError(loc: parentLoc) |
| 1251 | << "NYI: address point of a relative layout vtable" ; |
| 1252 | return {}; |
| 1253 | } |
| 1254 | |
| 1255 | // Like classic codegen, a vtable address point is inbounds, and only the |
| 1256 | // one vtable of the group it points into can be accessed through it. |
| 1257 | auto indices = |
| 1258 | globalAttr.getIndices().getAsValueRange<mlir::IntegerAttr>(); |
| 1259 | assert(llvm::range_size(indices) == 2 && |
| 1260 | "address point takes a vtable index and a slot index" ); |
| 1261 | auto vtableTy = mlir::cast<mlir::LLVM::LLVMArrayType>( |
| 1262 | Val: mlir::cast<mlir::LLVM::LLVMStructType>(Val&: eltTy) |
| 1263 | .getBody()[(*indices.begin()).getZExtValue()]); |
| 1264 | mlir::DataLayout layout(parentModule); |
| 1265 | int64_t slotSize = layout.getTypeSize(t: vtableTy.getElementType()); |
| 1266 | int64_t offset = (*std::next(x: indices.begin())).getSExtValue() * slotSize; |
| 1267 | int64_t size = vtableTy.getNumElements() * slotSize; |
| 1268 | unsigned indexBits = *layout.getTypeIndexBitwidth( |
| 1269 | t: mlir::cast<mlir::LLVM::LLVMPointerType>(Val: addrOp.getType())); |
| 1270 | gep.setNoWrapFlags(mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 1271 | |
| 1272 | gep.setInrangeAttr(mlir::LLVM::ConstantRangeAttr::get( |
| 1273 | context: rewriter.getContext(), |
| 1274 | lower: llvm::APInt(indexBits, -offset, /*isSigned=*/true), |
| 1275 | upper: llvm::APInt(indexBits, size - offset, /*isSigned=*/true))); |
| 1276 | } |
| 1277 | addrOp = gep; |
| 1278 | } |
| 1279 | |
| 1280 | return castGlobalAddrToType(addr: addrOp, attrType: globalAttr.getType(), sourceType, |
| 1281 | addrSpace: lowered.addrSpace); |
| 1282 | } |
| 1283 | |
| 1284 | mlir::Value CIRAttrToValue::visitCirAttr(cir::GlobalOffsetAttr globalAttr) { |
| 1285 | LoweredGlobalSymbol lowered = getAddressOfSymbol(symbol: globalAttr.getSymbol()); |
| 1286 | mlir::Value addrOp = lowered.addr; |
| 1287 | |
| 1288 | // A displacement in bytes is a GEP through i8, which is what classic |
| 1289 | // codegen emits for the offset of any constant lvalue. |
| 1290 | mlir::Value offset = mlir::LLVM::ConstantOp::create( |
| 1291 | builder&: rewriter, location: parentLoc, res: rewriter.getI64Type(), |
| 1292 | value: rewriter.getI64IntegerAttr(value: globalAttr.getOffset())); |
| 1293 | addrOp = mlir::LLVM::GEPOp::create( |
| 1294 | builder&: rewriter, location: parentLoc, resultType: addrOp.getType(), elementType: rewriter.getI8Type(), basePtr: addrOp, |
| 1295 | indices: mlir::ValueRange{offset}, noWrapFlags: mlir::LLVM::GEPNoWrapFlags::none); |
| 1296 | |
| 1297 | return castGlobalAddrToType(addr: addrOp, attrType: globalAttr.getType(), sourceType: lowered.sourceType, |
| 1298 | addrSpace: lowered.addrSpace); |
| 1299 | } |
| 1300 | |
| 1301 | // TypeInfoAttr visitor. |
| 1302 | mlir::Value CIRAttrToValue::visitCirAttr(cir::TypeInfoAttr typeInfoAttr) { |
| 1303 | mlir::Type llvmTy = converter->convertType(t: typeInfoAttr.getType()); |
| 1304 | mlir::Value result = mlir::LLVM::UndefOp::create(builder&: rewriter, location: parentLoc, res: llvmTy); |
| 1305 | |
| 1306 | for (auto [idx, elt] : llvm::enumerate(First: typeInfoAttr.getData())) { |
| 1307 | mlir::Value init = visit(attr: elt); |
| 1308 | if (!init) |
| 1309 | return {}; |
| 1310 | result = mlir::LLVM::InsertValueOp::create(builder&: rewriter, location: parentLoc, container: result, |
| 1311 | value: init, position: idx); |
| 1312 | } |
| 1313 | |
| 1314 | return result; |
| 1315 | } |
| 1316 | |
| 1317 | /// UndefAttr visitor. |
| 1318 | mlir::Value CIRAttrToValue::visitCirAttr(cir::UndefAttr undefAttr) { |
| 1319 | mlir::DataLayout layout(parentModule); |
| 1320 | return mlir::LLVM::UndefOp::create( |
| 1321 | builder&: rewriter, location: parentLoc, |
| 1322 | res: convertTypeForMemory(converter: *converter, dataLayout: layout, type: undefAttr.getType())); |
| 1323 | } |
| 1324 | |
| 1325 | /// PoisonAttr visitor. |
| 1326 | mlir::Value CIRAttrToValue::visitCirAttr(cir::PoisonAttr poisonAttr) { |
| 1327 | mlir::DataLayout layout(parentModule); |
| 1328 | return mlir::LLVM::PoisonOp::create( |
| 1329 | builder&: rewriter, location: parentLoc, |
| 1330 | res: convertTypeForMemory(converter: *converter, dataLayout: layout, type: poisonAttr.getType())); |
| 1331 | } |
| 1332 | |
| 1333 | // VTableAttr visitor. |
| 1334 | mlir::Value CIRAttrToValue::visitCirAttr(cir::VTableAttr vtableArr) { |
| 1335 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 1336 | if (usesRelativeVTables(mod: parentModule)) { |
| 1337 | mlir::emitError(loc: parentLoc) |
| 1338 | << "NYI: initializer of a relative layout vtable" ; |
| 1339 | return {}; |
| 1340 | } |
| 1341 | |
| 1342 | mlir::Type llvmTy = converter->convertType(t: vtableArr.getType()); |
| 1343 | mlir::Value result = mlir::LLVM::UndefOp::create(builder&: rewriter, location: parentLoc, res: llvmTy); |
| 1344 | |
| 1345 | for (auto [idx, elt] : llvm::enumerate(First: vtableArr.getData())) { |
| 1346 | mlir::Value init = visit(attr: elt); |
| 1347 | if (!init) |
| 1348 | return {}; |
| 1349 | result = mlir::LLVM::InsertValueOp::create(builder&: rewriter, location: parentLoc, container: result, |
| 1350 | value: init, position: idx); |
| 1351 | } |
| 1352 | |
| 1353 | return result; |
| 1354 | } |
| 1355 | |
| 1356 | /// ZeroAttr visitor. |
| 1357 | mlir::Value CIRAttrToValue::visitCirAttr(cir::ZeroAttr attr) { |
| 1358 | mlir::DataLayout layout(parentModule); |
| 1359 | return mlir::LLVM::ZeroOp::create( |
| 1360 | builder&: rewriter, location: parentLoc, |
| 1361 | res: convertTypeForMemory(converter: *converter, dataLayout: layout, type: attr.getType())); |
| 1362 | } |
| 1363 | |
| 1364 | // This class handles rewriting initializer attributes for types that do not |
| 1365 | // require region initialization. |
| 1366 | class GlobalInitAttrRewriter { |
| 1367 | public: |
| 1368 | GlobalInitAttrRewriter(mlir::Type type, mlir::DataLayout const &dataLayout, |
| 1369 | mlir::ConversionPatternRewriter &rewriter) |
| 1370 | : llvmType(type), dataLayout(dataLayout), rewriter(rewriter) {} |
| 1371 | |
| 1372 | mlir::Attribute visit(mlir::Attribute attr) { |
| 1373 | return llvm::TypeSwitch<mlir::Attribute, mlir::Attribute>(attr) |
| 1374 | .Case<cir::IntAttr, cir::FPAttr, cir::BoolAttr>( |
| 1375 | caseFn: [&](auto attrT) { return visitCirAttr(attrT); }) |
| 1376 | .Default(defaultFn: [&](auto attrT) { return mlir::Attribute(); }); |
| 1377 | } |
| 1378 | |
| 1379 | mlir::Attribute visitCirAttr(cir::IntAttr attr) { |
| 1380 | // A split-storage _BitInt's memory representation is a byte array (see |
| 1381 | // convertTypeForMemory), not a scalar, so its constant must be built as |
| 1382 | // raw bytes rather than a single sign/zero-extended IntegerAttr. |
| 1383 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val: attr.getType()); |
| 1384 | intTy && intTy.isBitInt()) |
| 1385 | return getBitIntStorageAttr(rewriter, attr, dataLayout); |
| 1386 | |
| 1387 | return rewriter.getIntegerAttr(type: llvmType, value: attr.getValue()); |
| 1388 | } |
| 1389 | |
| 1390 | mlir::Attribute visitCirAttr(cir::FPAttr attr) { |
| 1391 | return rewriter.getFloatAttr(type: llvmType, value: attr.getValue()); |
| 1392 | } |
| 1393 | |
| 1394 | mlir::Attribute visitCirAttr(cir::BoolAttr attr) { |
| 1395 | return rewriter.getBoolAttr(value: attr.getValue()); |
| 1396 | } |
| 1397 | |
| 1398 | private: |
| 1399 | mlir::Type llvmType; |
| 1400 | const mlir::DataLayout &dataLayout; |
| 1401 | mlir::ConversionPatternRewriter &rewriter; |
| 1402 | }; |
| 1403 | |
| 1404 | // This pass requires the CIR to be in a "flat" state. All blocks in each |
| 1405 | // function must belong to the parent region. Once scopes and control flow |
| 1406 | // are implemented in CIR, a pass will be run before this one to flatten |
| 1407 | // the CIR and get it into the state that this pass requires. |
| 1408 | struct ConvertCIRToLLVMPass |
| 1409 | : public mlir::PassWrapper<ConvertCIRToLLVMPass, |
| 1410 | mlir::OperationPass<mlir::ModuleOp>> { |
| 1411 | void getDependentDialects(mlir::DialectRegistry ®istry) const override { |
| 1412 | registry.insert<mlir::BuiltinDialect, mlir::DLTIDialect, |
| 1413 | mlir::LLVM::LLVMDialect, mlir::func::FuncDialect>(); |
| 1414 | } |
| 1415 | void runOnOperation() final; |
| 1416 | |
| 1417 | void processCIRAttrs(mlir::ModuleOp module); |
| 1418 | |
| 1419 | void resolveBlockAddressOp(LLVMBlockAddressInfo &blockInfoAddr); |
| 1420 | |
| 1421 | /// Collect (symbol_name, annotations, loc) from cir.func and cir.global ops |
| 1422 | /// before the conversion runs (the annotations attribute is dropped during |
| 1423 | /// FuncOp/GlobalOp lowering). |
| 1424 | void collectGlobalAnnotations(mlir::ModuleOp module); |
| 1425 | |
| 1426 | /// Emit @llvm.global.annotations and supporting string/args constants from |
| 1427 | /// the previously-collected annotations. Mirrors what OGCG produces. |
| 1428 | void buildGlobalAnnotationsVar(mlir::ModuleOp module, |
| 1429 | LLVMAnnotationGlobals &annotationGlobals); |
| 1430 | |
| 1431 | StringRef getDescription() const override { |
| 1432 | return "Convert the prepared CIR dialect module to LLVM dialect" ; |
| 1433 | } |
| 1434 | |
| 1435 | StringRef getArgument() const override { return "cir-flat-to-llvm" ; } |
| 1436 | |
| 1437 | private: |
| 1438 | /// One annotation entry collected pre-conversion. |
| 1439 | struct CollectedAnnotation { |
| 1440 | mlir::StringAttr symName; |
| 1441 | cir::AnnotationAttr annotation; |
| 1442 | mlir::Location loc; |
| 1443 | CollectedAnnotation(mlir::StringAttr symName, |
| 1444 | cir::AnnotationAttr annotation, mlir::Location loc) |
| 1445 | : symName(symName), annotation(annotation), loc(loc) {} |
| 1446 | }; |
| 1447 | llvm::SmallVector<CollectedAnnotation> collectedAnnotations; |
| 1448 | }; |
| 1449 | |
| 1450 | mlir::LogicalResult CIRToLLVMSignBitOpLowering::matchAndRewrite( |
| 1451 | cir::SignBitOp op, OpAdaptor adaptor, |
| 1452 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1453 | assert(!cir::MissingFeatures::isPPC_FP128Ty()); |
| 1454 | |
| 1455 | mlir::DataLayout layout(op->getParentOfType<mlir::ModuleOp>()); |
| 1456 | int width = layout.getTypeSizeInBits(t: op.getInput().getType()); |
| 1457 | if (auto longDoubleType = |
| 1458 | mlir::dyn_cast<cir::LongDoubleType>(Val: op.getInput().getType())) { |
| 1459 | if (mlir::isa<cir::FP80Type>(Val: longDoubleType.getUnderlying())) { |
| 1460 | // If the underlying type of LongDouble is FP80Type, |
| 1461 | // DataLayout::getTypeSizeInBits returns 128. |
| 1462 | // See https://github.com/llvm/clangir/issues/1057. |
| 1463 | // Set the width to 80 manually. |
| 1464 | width = 80; |
| 1465 | } |
| 1466 | } |
| 1467 | mlir::Type intTy = mlir::IntegerType::get(context: rewriter.getContext(), width); |
| 1468 | auto bitcast = mlir::LLVM::BitcastOp::create(builder&: rewriter, location: op->getLoc(), res: intTy, |
| 1469 | arg: adaptor.getInput()); |
| 1470 | |
| 1471 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op->getLoc(), type: intTy, value: 0); |
| 1472 | auto cmpResult = mlir::LLVM::ICmpOp::create(builder&: rewriter, location: op.getLoc(), |
| 1473 | predicate: mlir::LLVM::ICmpPredicate::slt, |
| 1474 | lhs: bitcast.getResult(), rhs: zero); |
| 1475 | rewriter.replaceOp(op, newOp: cmpResult); |
| 1476 | return mlir::success(); |
| 1477 | } |
| 1478 | |
| 1479 | mlir::LogicalResult CIRToLLVMAssumeOpLowering::matchAndRewrite( |
| 1480 | cir::AssumeOp op, OpAdaptor adaptor, |
| 1481 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1482 | mlir::Value cond = adaptor.getPredicate(); |
| 1483 | if (op.getBundleKind() == cir::AssumeBundleKind::None) { |
| 1484 | rewriter.replaceOpWithNewOp<mlir::LLVM::AssumeOp>(op, args&: cond); |
| 1485 | return mlir::success(); |
| 1486 | } |
| 1487 | |
| 1488 | llvm::StringRef tag = cir::stringifyAssumeBundleKind(op.getBundleKind()); |
| 1489 | rewriter.replaceOpWithNewOp<mlir::LLVM::AssumeOp>(op, args&: cond, args&: tag, |
| 1490 | args: adaptor.getBundleArgs()); |
| 1491 | return mlir::success(); |
| 1492 | } |
| 1493 | |
| 1494 | static mlir::LLVM::AtomicOrdering |
| 1495 | getLLVMMemOrder(std::optional<cir::MemOrder> memorder) { |
| 1496 | if (!memorder) |
| 1497 | return mlir::LLVM::AtomicOrdering::not_atomic; |
| 1498 | switch (*memorder) { |
| 1499 | case cir::MemOrder::Relaxed: |
| 1500 | return mlir::LLVM::AtomicOrdering::monotonic; |
| 1501 | case cir::MemOrder::Consume: |
| 1502 | case cir::MemOrder::Acquire: |
| 1503 | return mlir::LLVM::AtomicOrdering::acquire; |
| 1504 | case cir::MemOrder::Release: |
| 1505 | return mlir::LLVM::AtomicOrdering::release; |
| 1506 | case cir::MemOrder::AcquireRelease: |
| 1507 | return mlir::LLVM::AtomicOrdering::acq_rel; |
| 1508 | case cir::MemOrder::SequentiallyConsistent: |
| 1509 | return mlir::LLVM::AtomicOrdering::seq_cst; |
| 1510 | } |
| 1511 | llvm_unreachable("unknown memory order" ); |
| 1512 | } |
| 1513 | |
| 1514 | static llvm::AtomicScope getAtomicScope(cir::SyncScopeKind syncScope) { |
| 1515 | switch (syncScope) { |
| 1516 | case cir::SyncScopeKind::SingleThread: |
| 1517 | case cir::SyncScopeKind::HIPSingleThread: |
| 1518 | return llvm::AtomicScope::Single; |
| 1519 | case cir::SyncScopeKind::Wavefront: |
| 1520 | case cir::SyncScopeKind::HIPWavefront: |
| 1521 | case cir::SyncScopeKind::OpenCLSubGroup: |
| 1522 | return llvm::AtomicScope::Wavefront; |
| 1523 | case cir::SyncScopeKind::Workgroup: |
| 1524 | case cir::SyncScopeKind::HIPWorkgroup: |
| 1525 | case cir::SyncScopeKind::OpenCLWorkGroup: |
| 1526 | return llvm::AtomicScope::Workgroup; |
| 1527 | case cir::SyncScopeKind::Cluster: |
| 1528 | case cir::SyncScopeKind::HIPCluster: |
| 1529 | return llvm::AtomicScope::Cluster; |
| 1530 | case cir::SyncScopeKind::Device: |
| 1531 | case cir::SyncScopeKind::HIPAgent: |
| 1532 | case cir::SyncScopeKind::OpenCLDevice: |
| 1533 | return llvm::AtomicScope::Device; |
| 1534 | case cir::SyncScopeKind::System: |
| 1535 | case cir::SyncScopeKind::HIPSystem: |
| 1536 | case cir::SyncScopeKind::OpenCLAllSVMDevices: |
| 1537 | return llvm::AtomicScope::System; |
| 1538 | } |
| 1539 | llvm_unreachable("unknown sync scope" ); |
| 1540 | } |
| 1541 | |
| 1542 | static llvm::StringRef getLLVMSyncScope(cir::SyncScopeKind syncScope, |
| 1543 | mlir::ModuleOp moduleOp) { |
| 1544 | llvm::AtomicScope scope = getAtomicScope(syncScope); |
| 1545 | if (auto triple = moduleOp->getAttrOfType<mlir::StringAttr>( |
| 1546 | name: cir::CIRDialect::getTripleAttrName())) |
| 1547 | if (std::optional<llvm::StringRef> name = llvm::getAtomicScopeIRString( |
| 1548 | T: llvm::Triple(triple.getValue()), S: scope)) |
| 1549 | return *name; |
| 1550 | return scope == llvm::AtomicScope::Single ? "singlethread" : "" ; |
| 1551 | } |
| 1552 | |
| 1553 | static std::optional<llvm::StringRef> |
| 1554 | getLLVMSyncScope(std::optional<cir::SyncScopeKind> syncScope, |
| 1555 | mlir::ModuleOp moduleOp) { |
| 1556 | if (syncScope.has_value()) |
| 1557 | return getLLVMSyncScope(syncScope: *syncScope, moduleOp); |
| 1558 | return std::nullopt; |
| 1559 | } |
| 1560 | |
| 1561 | static void forwardAMDGPUNoPrivateAddrSpace(mlir::Operation *op, |
| 1562 | mlir::Operation *llvmOp, |
| 1563 | mlir::Value llvmPtr) { |
| 1564 | llvm::StringRef name = cir::CIRDialect::getAMDGPUNoPrivateAddrSpaceAttrName(); |
| 1565 | auto ptrTy = mlir::cast<mlir::LLVM::LLVMPointerType>(Val: llvmPtr.getType()); |
| 1566 | mlir::Attribute a = op->getAttr(name); |
| 1567 | if (a && ptrTy.getAddressSpace() == llvm::AMDGPUAS::FLAT_ADDRESS) |
| 1568 | llvmOp->setAttr(name, value: a); |
| 1569 | } |
| 1570 | |
| 1571 | // CIRGen decides the metadata for a C++/HIP atomic from the atomic options |
| 1572 | // in effect, so those markers are simply carried across. |
| 1573 | static void forwardAMDGPUAtomicOptionMarkers(mlir::Operation *op, |
| 1574 | mlir::Operation *llvmOp) { |
| 1575 | for (llvm::StringRef marker : |
| 1576 | {cir::CIRDialect::getAMDGPUNoFineGrainedMemoryAttrName(), |
| 1577 | cir::CIRDialect::getAMDGPUNoRemoteMemoryAttrName(), |
| 1578 | cir::CIRDialect::getAMDGPUIgnoreDenormalModeAttrName()}) |
| 1579 | if (mlir::Attribute a = op->getAttr(name: marker)) |
| 1580 | llvmOp->setAttr(name: marker, value: a); |
| 1581 | } |
| 1582 | |
| 1583 | mlir::LogicalResult CIRToLLVMAtomicCmpXchgOpLowering::matchAndRewrite( |
| 1584 | cir::AtomicCmpXchgOp op, OpAdaptor adaptor, |
| 1585 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1586 | mlir::Value expected = adaptor.getExpected(); |
| 1587 | mlir::Value desired = adaptor.getDesired(); |
| 1588 | |
| 1589 | auto cmpxchg = mlir::LLVM::AtomicCmpXchgOp::create( |
| 1590 | builder&: rewriter, location: op.getLoc(), ptr: adaptor.getPtr(), cmp: expected, val: desired, |
| 1591 | successOrdering: getLLVMMemOrder(memorder: adaptor.getSuccOrder()), |
| 1592 | failureOrdering: getLLVMMemOrder(memorder: adaptor.getFailOrder()), |
| 1593 | syncscope: getLLVMSyncScope(syncScope: op.getSyncScope(), |
| 1594 | moduleOp: op->getParentOfType<mlir::ModuleOp>())); |
| 1595 | |
| 1596 | cmpxchg.setAlignment(adaptor.getAlignment()); |
| 1597 | cmpxchg.setWeak(adaptor.getWeak()); |
| 1598 | cmpxchg.setVolatile_(adaptor.getIsVolatile()); |
| 1599 | forwardAMDGPUNoPrivateAddrSpace(op, llvmOp: cmpxchg, llvmPtr: adaptor.getPtr()); |
| 1600 | |
| 1601 | // Check result and apply stores accordingly. |
| 1602 | auto old = mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: op.getLoc(), |
| 1603 | container: cmpxchg.getResult(), position: 0); |
| 1604 | auto cmp = mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: op.getLoc(), |
| 1605 | container: cmpxchg.getResult(), position: 1); |
| 1606 | |
| 1607 | rewriter.replaceOp(op, newValues: {old, cmp}); |
| 1608 | return mlir::success(); |
| 1609 | } |
| 1610 | |
| 1611 | mlir::LogicalResult CIRToLLVMAtomicXchgOpLowering::matchAndRewrite( |
| 1612 | cir::AtomicXchgOp op, OpAdaptor adaptor, |
| 1613 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1614 | assert(!cir::MissingFeatures::atomicSyncScopeID()); |
| 1615 | mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(memorder: adaptor.getMemOrder()); |
| 1616 | llvm::StringRef llvmSyncScope = getLLVMSyncScope( |
| 1617 | syncScope: op.getSyncScope(), moduleOp: op->getParentOfType<mlir::ModuleOp>()); |
| 1618 | // Create the new op first: the forward helpers read `op`, which the |
| 1619 | // replace below erases immediately. |
| 1620 | auto xchg = mlir::LLVM::AtomicRMWOp::create( |
| 1621 | builder&: rewriter, location: op.getLoc(), binOp: mlir::LLVM::AtomicBinOp::xchg, ptr: adaptor.getPtr(), |
| 1622 | val: adaptor.getVal(), ordering: llvmOrder, syncscope: llvmSyncScope, /*alignment=*/0, |
| 1623 | isVolatile: op.getIsVolatile()); |
| 1624 | forwardAMDGPUAtomicOptionMarkers(op, llvmOp: xchg); |
| 1625 | forwardAMDGPUNoPrivateAddrSpace(op, llvmOp: xchg, llvmPtr: adaptor.getPtr()); |
| 1626 | rewriter.replaceOp(op, newOp: xchg); |
| 1627 | return mlir::success(); |
| 1628 | } |
| 1629 | |
| 1630 | mlir::LogicalResult CIRToLLVMAtomicTestAndSetOpLowering::matchAndRewrite( |
| 1631 | cir::AtomicTestAndSetOp op, OpAdaptor adaptor, |
| 1632 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1633 | assert(!cir::MissingFeatures::atomicSyncScopeID()); |
| 1634 | |
| 1635 | mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(memorder: op.getMemOrder()); |
| 1636 | |
| 1637 | auto one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 1638 | type: rewriter.getI8Type(), value: 1); |
| 1639 | auto rmw = mlir::LLVM::AtomicRMWOp::create( |
| 1640 | builder&: rewriter, location: op.getLoc(), binOp: mlir::LLVM::AtomicBinOp::xchg, ptr: adaptor.getPtr(), |
| 1641 | val: one, ordering: llvmOrder, /*syncscope=*/llvm::StringRef(), |
| 1642 | alignment: adaptor.getAlignment().value_or(u: 0), isVolatile: op.getIsVolatile()); |
| 1643 | |
| 1644 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 1645 | type: rewriter.getI8Type(), value: 0); |
| 1646 | auto cmp = mlir::LLVM::ICmpOp::create( |
| 1647 | builder&: rewriter, location: op.getLoc(), predicate: mlir::LLVM::ICmpPredicate::ne, lhs: rmw, rhs: zero); |
| 1648 | |
| 1649 | rewriter.replaceOp(op, newOp: cmp); |
| 1650 | return mlir::success(); |
| 1651 | } |
| 1652 | |
| 1653 | mlir::LogicalResult CIRToLLVMAtomicClearOpLowering::matchAndRewrite( |
| 1654 | cir::AtomicClearOp op, OpAdaptor adaptor, |
| 1655 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1656 | assert(!cir::MissingFeatures::atomicSyncScopeID()); |
| 1657 | |
| 1658 | mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(memorder: op.getMemOrder()); |
| 1659 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 1660 | type: rewriter.getI8Type(), value: 0); |
| 1661 | auto store = mlir::LLVM::StoreOp::create( |
| 1662 | builder&: rewriter, location: op.getLoc(), value: zero, addr: adaptor.getPtr(), |
| 1663 | alignment: adaptor.getAlignment().value_or(u: 0), isVolatile: op.getIsVolatile(), |
| 1664 | /*isNonTemporal=*/false, /*isInvariantGroup=*/false, ordering: llvmOrder); |
| 1665 | |
| 1666 | rewriter.replaceOp(op, newOp: store); |
| 1667 | return mlir::success(); |
| 1668 | } |
| 1669 | |
| 1670 | mlir::LogicalResult CIRToLLVMAtomicFenceOpLowering::matchAndRewrite( |
| 1671 | cir::AtomicFenceOp op, OpAdaptor adaptor, |
| 1672 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1673 | mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(memorder: adaptor.getOrdering()); |
| 1674 | |
| 1675 | auto fence = mlir::LLVM::FenceOp::create(builder&: rewriter, location: op.getLoc(), ordering: llvmOrder); |
| 1676 | fence.setSyncscope(getLLVMSyncScope(syncScope: op.getSyncscope(), |
| 1677 | moduleOp: op->getParentOfType<mlir::ModuleOp>())); |
| 1678 | if (mlir::Attribute mmra = |
| 1679 | op->getAttr(name: mlir::LLVM::LLVMDialect::getMmraAttrName())) |
| 1680 | fence->setAttr(name: mlir::LLVM::LLVMDialect::getMmraAttrName(), value: mmra); |
| 1681 | |
| 1682 | rewriter.replaceOp(op, newOp: fence); |
| 1683 | |
| 1684 | return mlir::success(); |
| 1685 | } |
| 1686 | |
| 1687 | static mlir::LLVM::AtomicBinOp |
| 1688 | getLLVMAtomicBinOp(cir::AtomicFetchKind k, bool isInt, bool isSignedInt) { |
| 1689 | switch (k) { |
| 1690 | case cir::AtomicFetchKind::Add: |
| 1691 | return isInt ? mlir::LLVM::AtomicBinOp::add : mlir::LLVM::AtomicBinOp::fadd; |
| 1692 | case cir::AtomicFetchKind::Sub: |
| 1693 | return isInt ? mlir::LLVM::AtomicBinOp::sub : mlir::LLVM::AtomicBinOp::fsub; |
| 1694 | case cir::AtomicFetchKind::And: |
| 1695 | return mlir::LLVM::AtomicBinOp::_and; |
| 1696 | case cir::AtomicFetchKind::Xor: |
| 1697 | return mlir::LLVM::AtomicBinOp::_xor; |
| 1698 | case cir::AtomicFetchKind::Or: |
| 1699 | return mlir::LLVM::AtomicBinOp::_or; |
| 1700 | case cir::AtomicFetchKind::Nand: |
| 1701 | return mlir::LLVM::AtomicBinOp::nand; |
| 1702 | case cir::AtomicFetchKind::Max: { |
| 1703 | if (!isInt) |
| 1704 | return mlir::LLVM::AtomicBinOp::fmax; |
| 1705 | return isSignedInt ? mlir::LLVM::AtomicBinOp::max |
| 1706 | : mlir::LLVM::AtomicBinOp::umax; |
| 1707 | } |
| 1708 | case cir::AtomicFetchKind::Min: { |
| 1709 | if (!isInt) |
| 1710 | return mlir::LLVM::AtomicBinOp::fmin; |
| 1711 | return isSignedInt ? mlir::LLVM::AtomicBinOp::min |
| 1712 | : mlir::LLVM::AtomicBinOp::umin; |
| 1713 | } |
| 1714 | case cir::AtomicFetchKind::UIncWrap: |
| 1715 | return mlir::LLVM::AtomicBinOp::uinc_wrap; |
| 1716 | case cir::AtomicFetchKind::UDecWrap: |
| 1717 | return mlir::LLVM::AtomicBinOp::udec_wrap; |
| 1718 | case cir::AtomicFetchKind::Maximum: |
| 1719 | return mlir::LLVM::AtomicBinOp::fmaximum; |
| 1720 | case cir::AtomicFetchKind::Minimum: |
| 1721 | return mlir::LLVM::AtomicBinOp::fminimum; |
| 1722 | case cir::AtomicFetchKind::MaximumNum: |
| 1723 | return mlir::LLVM::AtomicBinOp::fmaximumnum; |
| 1724 | case cir::AtomicFetchKind::MinimumNum: |
| 1725 | return mlir::LLVM::AtomicBinOp::fminimumnum; |
| 1726 | } |
| 1727 | llvm_unreachable("Unknown atomic fetch opcode" ); |
| 1728 | } |
| 1729 | |
| 1730 | static llvm::StringLiteral getLLVMBinopForPostAtomic(cir::AtomicFetchKind k, |
| 1731 | bool isInt) { |
| 1732 | switch (k) { |
| 1733 | case cir::AtomicFetchKind::Add: |
| 1734 | return isInt ? mlir::LLVM::AddOp::getOperationName() |
| 1735 | : mlir::LLVM::FAddOp::getOperationName(); |
| 1736 | case cir::AtomicFetchKind::Sub: |
| 1737 | return isInt ? mlir::LLVM::SubOp::getOperationName() |
| 1738 | : mlir::LLVM::FSubOp::getOperationName(); |
| 1739 | case cir::AtomicFetchKind::And: |
| 1740 | return mlir::LLVM::AndOp::getOperationName(); |
| 1741 | case cir::AtomicFetchKind::Xor: |
| 1742 | return mlir::LLVM::XOrOp::getOperationName(); |
| 1743 | case cir::AtomicFetchKind::Or: |
| 1744 | return mlir::LLVM::OrOp::getOperationName(); |
| 1745 | case cir::AtomicFetchKind::Nand: |
| 1746 | // There's no nand binop in LLVM, this is later fixed with a not. |
| 1747 | return mlir::LLVM::AndOp::getOperationName(); |
| 1748 | case cir::AtomicFetchKind::Max: |
| 1749 | case cir::AtomicFetchKind::Min: |
| 1750 | llvm_unreachable("handled in buildMinMaxPostOp" ); |
| 1751 | case cir::AtomicFetchKind::UIncWrap: |
| 1752 | case cir::AtomicFetchKind::UDecWrap: |
| 1753 | case cir::AtomicFetchKind::Maximum: |
| 1754 | case cir::AtomicFetchKind::Minimum: |
| 1755 | case cir::AtomicFetchKind::MaximumNum: |
| 1756 | case cir::AtomicFetchKind::MinimumNum: |
| 1757 | llvm_unreachable("uinc_wrap, udec_wrap, maximum, minimum, maximum_num, and " |
| 1758 | "minimum_num are always fetch_first" ); |
| 1759 | } |
| 1760 | llvm_unreachable("Unknown atomic fetch opcode" ); |
| 1761 | } |
| 1762 | |
| 1763 | mlir::Value CIRToLLVMAtomicFetchOpLowering::buildPostOp( |
| 1764 | cir::AtomicFetchOp op, OpAdaptor adaptor, |
| 1765 | mlir::ConversionPatternRewriter &rewriter, mlir::Value rmwVal, |
| 1766 | bool isInt) const { |
| 1767 | SmallVector<mlir::Value> atomicOperands = {rmwVal, adaptor.getVal()}; |
| 1768 | SmallVector<mlir::Type> atomicResTys = {rmwVal.getType()}; |
| 1769 | return rewriter |
| 1770 | .create(loc: op.getLoc(), |
| 1771 | opName: rewriter.getStringAttr( |
| 1772 | bytes: getLLVMBinopForPostAtomic(k: op.getBinop(), isInt)), |
| 1773 | operands: atomicOperands, types: atomicResTys, attributes: {}) |
| 1774 | ->getResult(idx: 0); |
| 1775 | } |
| 1776 | |
| 1777 | mlir::Value CIRToLLVMAtomicFetchOpLowering::buildMinMaxPostOp( |
| 1778 | cir::AtomicFetchOp op, OpAdaptor adaptor, |
| 1779 | mlir::ConversionPatternRewriter &rewriter, mlir::Value rmwVal, bool isInt, |
| 1780 | bool isSigned) const { |
| 1781 | mlir::Location loc = op.getLoc(); |
| 1782 | |
| 1783 | if (!isInt) { |
| 1784 | if (op.getBinop() == cir::AtomicFetchKind::Max) |
| 1785 | return mlir::LLVM::MaxNumOp::create(builder&: rewriter, location: loc, a: rmwVal, |
| 1786 | b: adaptor.getVal()); |
| 1787 | return mlir::LLVM::MinNumOp::create(builder&: rewriter, location: loc, a: rmwVal, |
| 1788 | b: adaptor.getVal()); |
| 1789 | } |
| 1790 | |
| 1791 | mlir::LLVM::ICmpPredicate pred; |
| 1792 | if (op.getBinop() == cir::AtomicFetchKind::Max) { |
| 1793 | pred = isSigned ? mlir::LLVM::ICmpPredicate::sgt |
| 1794 | : mlir::LLVM::ICmpPredicate::ugt; |
| 1795 | } else { // Min |
| 1796 | pred = isSigned ? mlir::LLVM::ICmpPredicate::slt |
| 1797 | : mlir::LLVM::ICmpPredicate::ult; |
| 1798 | } |
| 1799 | mlir::Value cmp = mlir::LLVM::ICmpOp::create( |
| 1800 | builder&: rewriter, location: loc, |
| 1801 | predicate: mlir::LLVM::ICmpPredicateAttr::get(context: rewriter.getContext(), val: pred), lhs: rmwVal, |
| 1802 | rhs: adaptor.getVal()); |
| 1803 | return mlir::LLVM::SelectOp::create(builder&: rewriter, location: loc, condition: cmp, trueValue: rmwVal, |
| 1804 | falseValue: adaptor.getVal()); |
| 1805 | } |
| 1806 | |
| 1807 | mlir::LogicalResult CIRToLLVMAtomicFetchOpLowering::matchAndRewrite( |
| 1808 | cir::AtomicFetchOp op, OpAdaptor adaptor, |
| 1809 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1810 | bool isInt = false; |
| 1811 | bool isSignedInt = false; |
| 1812 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val: op.getVal().getType())) { |
| 1813 | isInt = true; |
| 1814 | isSignedInt = intTy.isSigned(); |
| 1815 | } else if (cir::isFPOrVectorOfFPType(type: op.getVal().getType())) { |
| 1816 | isInt = false; |
| 1817 | } else { |
| 1818 | return op.emitError() << "Unsupported type: " << op.getVal().getType(); |
| 1819 | } |
| 1820 | |
| 1821 | mlir::LLVM::AtomicOrdering llvmOrder = getLLVMMemOrder(memorder: op.getMemOrder()); |
| 1822 | llvm::StringRef llvmSyncScope = getLLVMSyncScope( |
| 1823 | syncScope: op.getSyncScope(), moduleOp: op->getParentOfType<mlir::ModuleOp>()); |
| 1824 | mlir::LLVM::AtomicBinOp llvmBinOp = |
| 1825 | getLLVMAtomicBinOp(k: op.getBinop(), isInt, isSignedInt); |
| 1826 | auto rmwVal = mlir::LLVM::AtomicRMWOp::create( |
| 1827 | builder&: rewriter, location: op.getLoc(), binOp: llvmBinOp, ptr: adaptor.getPtr(), val: adaptor.getVal(), |
| 1828 | ordering: llvmOrder, syncscope: llvmSyncScope, /*alignment=*/0, isVolatile: op.getIsVolatile()); |
| 1829 | |
| 1830 | forwardAMDGPUAtomicOptionMarkers(op, llvmOp: rmwVal); |
| 1831 | forwardAMDGPUNoPrivateAddrSpace(op, llvmOp: rmwVal, llvmPtr: adaptor.getPtr()); |
| 1832 | |
| 1833 | // The AMDGPU raw hardware atomic builtins need metadata for the backend to |
| 1834 | // select the native instruction. LDS atomics are always native, so the |
| 1835 | // metadata is only needed for the global and flat address spaces. |
| 1836 | if (op->hasAttr(name: cir::CIRDialect::getAMDGPURawAtomicAttrName())) { |
| 1837 | auto ptrTy = |
| 1838 | mlir::cast<mlir::LLVM::LLVMPointerType>(Val: adaptor.getPtr().getType()); |
| 1839 | if (ptrTy.getAddressSpace() != llvm::AMDGPUAS::LOCAL_ADDRESS) { |
| 1840 | mlir::UnitAttr unit = rewriter.getUnitAttr(); |
| 1841 | rmwVal->setAttr(name: cir::CIRDialect::getAMDGPUNoFineGrainedMemoryAttrName(), |
| 1842 | value: unit); |
| 1843 | // Denormal flushing only matters for a float add. |
| 1844 | if (llvmBinOp == mlir::LLVM::AtomicBinOp::fadd && |
| 1845 | mlir::isa<cir::SingleType>(Val: op.getVal().getType())) |
| 1846 | rmwVal->setAttr(name: cir::CIRDialect::getAMDGPUIgnoreDenormalModeAttrName(), |
| 1847 | value: unit); |
| 1848 | } |
| 1849 | } |
| 1850 | |
| 1851 | mlir::Value result = rmwVal.getResult(); |
| 1852 | if (!op.getFetchFirst()) { |
| 1853 | if (op.getBinop() == cir::AtomicFetchKind::Max || |
| 1854 | op.getBinop() == cir::AtomicFetchKind::Min) |
| 1855 | result = buildMinMaxPostOp(op, adaptor, rewriter, rmwVal: rmwVal.getRes(), isInt, |
| 1856 | isSigned: isSignedInt); |
| 1857 | else |
| 1858 | result = buildPostOp(op, adaptor, rewriter, rmwVal: rmwVal.getRes(), isInt); |
| 1859 | |
| 1860 | // Compensate lack of nand binop in LLVM IR. |
| 1861 | if (op.getBinop() == cir::AtomicFetchKind::Nand) { |
| 1862 | auto negOne = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 1863 | type: result.getType(), value: -1); |
| 1864 | result = mlir::LLVM::XOrOp::create(builder&: rewriter, location: op.getLoc(), lhs: result, rhs: negOne); |
| 1865 | } |
| 1866 | } |
| 1867 | |
| 1868 | rewriter.replaceOp(op, newValues: result); |
| 1869 | return mlir::success(); |
| 1870 | } |
| 1871 | |
| 1872 | mlir::LogicalResult CIRToLLVMBitClrsbOpLowering::matchAndRewrite( |
| 1873 | cir::BitClrsbOp op, OpAdaptor adaptor, |
| 1874 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1875 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 1876 | type: adaptor.getInput().getType(), value: 0); |
| 1877 | auto isNeg = mlir::LLVM::ICmpOp::create( |
| 1878 | builder&: rewriter, location: op.getLoc(), |
| 1879 | predicate: mlir::LLVM::ICmpPredicateAttr::get(context: rewriter.getContext(), |
| 1880 | val: mlir::LLVM::ICmpPredicate::slt), |
| 1881 | lhs: adaptor.getInput(), rhs: zero); |
| 1882 | |
| 1883 | auto negOne = mlir::LLVM::ConstantOp::create( |
| 1884 | builder&: rewriter, location: op.getLoc(), type: adaptor.getInput().getType(), value: -1); |
| 1885 | auto flipped = mlir::LLVM::XOrOp::create(builder&: rewriter, location: op.getLoc(), |
| 1886 | lhs: adaptor.getInput(), rhs: negOne); |
| 1887 | |
| 1888 | auto select = mlir::LLVM::SelectOp::create(builder&: rewriter, location: op.getLoc(), condition: isNeg, |
| 1889 | trueValue: flipped, falseValue: adaptor.getInput()); |
| 1890 | |
| 1891 | auto resTy = getTypeConverter()->convertType(t: op.getType()); |
| 1892 | auto clz = mlir::LLVM::CountLeadingZerosOp::create( |
| 1893 | builder&: rewriter, location: op.getLoc(), res: resTy, in: select, /*is_zero_poison=*/false); |
| 1894 | |
| 1895 | auto one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), type: resTy, value: 1); |
| 1896 | auto res = mlir::LLVM::SubOp::create(builder&: rewriter, location: op.getLoc(), lhs: clz, rhs: one, |
| 1897 | overflowFlags: mlir::LLVM::IntegerOverflowFlags::nuw); |
| 1898 | rewriter.replaceOp(op, newOp: res); |
| 1899 | |
| 1900 | return mlir::LogicalResult::success(); |
| 1901 | } |
| 1902 | |
| 1903 | mlir::LogicalResult CIRToLLVMBitFfsOpLowering::matchAndRewrite( |
| 1904 | cir::BitFfsOp op, OpAdaptor adaptor, |
| 1905 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1906 | auto resTy = getTypeConverter()->convertType(t: op.getType()); |
| 1907 | auto ctz = mlir::LLVM::CountTrailingZerosOp::create(builder&: rewriter, location: op.getLoc(), |
| 1908 | res: resTy, in: adaptor.getInput(), |
| 1909 | /*is_zero_poison=*/true); |
| 1910 | |
| 1911 | auto one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), type: resTy, value: 1); |
| 1912 | auto ctzAddOne = mlir::LLVM::AddOp::create(builder&: rewriter, location: op.getLoc(), lhs: ctz, rhs: one); |
| 1913 | |
| 1914 | auto zeroInputTy = mlir::LLVM::ConstantOp::create( |
| 1915 | builder&: rewriter, location: op.getLoc(), type: adaptor.getInput().getType(), value: 0); |
| 1916 | auto isZero = mlir::LLVM::ICmpOp::create( |
| 1917 | builder&: rewriter, location: op.getLoc(), |
| 1918 | predicate: mlir::LLVM::ICmpPredicateAttr::get(context: rewriter.getContext(), |
| 1919 | val: mlir::LLVM::ICmpPredicate::eq), |
| 1920 | lhs: adaptor.getInput(), rhs: zeroInputTy); |
| 1921 | |
| 1922 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), type: resTy, value: 0); |
| 1923 | auto res = mlir::LLVM::SelectOp::create(builder&: rewriter, location: op.getLoc(), condition: isZero, trueValue: zero, |
| 1924 | falseValue: ctzAddOne); |
| 1925 | rewriter.replaceOp(op, newOp: res); |
| 1926 | |
| 1927 | return mlir::LogicalResult::success(); |
| 1928 | } |
| 1929 | |
| 1930 | mlir::LogicalResult CIRToLLVMBitParityOpLowering::matchAndRewrite( |
| 1931 | cir::BitParityOp op, OpAdaptor adaptor, |
| 1932 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1933 | auto resTy = getTypeConverter()->convertType(t: op.getType()); |
| 1934 | auto popcnt = mlir::LLVM::CtPopOp::create(builder&: rewriter, location: op.getLoc(), res: resTy, |
| 1935 | in: adaptor.getInput()); |
| 1936 | |
| 1937 | auto one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), type: resTy, value: 1); |
| 1938 | auto popcntMod2 = |
| 1939 | mlir::LLVM::AndOp::create(builder&: rewriter, location: op.getLoc(), lhs: popcnt, rhs: one); |
| 1940 | rewriter.replaceOp(op, newOp: popcntMod2); |
| 1941 | |
| 1942 | return mlir::LogicalResult::success(); |
| 1943 | } |
| 1944 | |
| 1945 | mlir::LogicalResult CIRToLLVMBrCondOpLowering::matchAndRewrite( |
| 1946 | cir::BrCondOp brOp, OpAdaptor adaptor, |
| 1947 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1948 | // When ZExtOp is implemented, we'll need to check if the condition is a |
| 1949 | // ZExtOp and if so, delete it if it has a single use. |
| 1950 | assert(!cir::MissingFeatures::zextOp()); |
| 1951 | |
| 1952 | mlir::Value i1Condition = adaptor.getCond(); |
| 1953 | |
| 1954 | rewriter.replaceOpWithNewOp<mlir::LLVM::CondBrOp>( |
| 1955 | op: brOp, args&: i1Condition, args: brOp.getDestTrue(), args: adaptor.getDestOperandsTrue(), |
| 1956 | args: brOp.getDestFalse(), args: adaptor.getDestOperandsFalse()); |
| 1957 | |
| 1958 | return mlir::success(); |
| 1959 | } |
| 1960 | |
| 1961 | mlir::Type CIRToLLVMCastOpLowering::convertTy(mlir::Type ty) const { |
| 1962 | return getTypeConverter()->convertType(t: ty); |
| 1963 | } |
| 1964 | |
| 1965 | mlir::LogicalResult CIRToLLVMCastOpLowering::matchAndRewrite( |
| 1966 | cir::CastOp castOp, OpAdaptor adaptor, |
| 1967 | mlir::ConversionPatternRewriter &rewriter) const { |
| 1968 | // For arithmetic conversions, LLVM IR uses the same instruction to convert |
| 1969 | // both individual scalars and entire vectors. This lowering pass handles |
| 1970 | // both situations. |
| 1971 | |
| 1972 | if (!getTypeConverter()->convertType(t: castOp.getType())) |
| 1973 | return emitTypeConversionError(op: castOp, type: castOp.getType()); |
| 1974 | |
| 1975 | switch (castOp.getKind()) { |
| 1976 | case cir::CastKind::array_to_ptrdecay: { |
| 1977 | const auto ptrTy = mlir::cast<cir::PointerType>(Val: castOp.getType()); |
| 1978 | mlir::Value sourceValue = adaptor.getSrc(); |
| 1979 | mlir::Type targetType = convertTy(ty: ptrTy); |
| 1980 | mlir::Type elementTy = convertTypeForMemory(converter: *getTypeConverter(), dataLayout, |
| 1981 | type: ptrTy.getPointee()); |
| 1982 | if (!elementTy) |
| 1983 | return emitTypeConversionError(op: castOp, type: ptrTy.getPointee()); |
| 1984 | llvm::SmallVector<mlir::LLVM::GEPArg> offset{0}; |
| 1985 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 1986 | op: castOp, args&: targetType, args&: elementTy, args&: sourceValue, args&: offset); |
| 1987 | break; |
| 1988 | } |
| 1989 | case cir::CastKind::int_to_bool: { |
| 1990 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 1991 | // getZeroAttr yields a splat for vector source types so this also |
| 1992 | // handles element-wise int-to-bool conversions (e.g. an ext_vector |
| 1993 | // __builtin_convertvector to bool). |
| 1994 | mlir::Value zeroInt = mlir::LLVM::ConstantOp::create( |
| 1995 | builder&: rewriter, location: castOp.getLoc(), res: llvmSrcVal.getType(), |
| 1996 | value: rewriter.getZeroAttr(type: llvmSrcVal.getType())); |
| 1997 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
| 1998 | op: castOp, args: mlir::LLVM::ICmpPredicate::ne, args&: llvmSrcVal, args&: zeroInt); |
| 1999 | break; |
| 2000 | } |
| 2001 | case cir::CastKind::integral: { |
| 2002 | mlir::Type srcType = castOp.getSrc().getType(); |
| 2003 | mlir::Type dstType = castOp.getType(); |
| 2004 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2005 | mlir::Type llvmDstType = getTypeConverter()->convertType(t: dstType); |
| 2006 | cir::IntType srcIntType = |
| 2007 | mlir::cast<cir::IntType>(Val: elementTypeIfVectorOrMatrix(type: srcType)); |
| 2008 | cir::IntType dstIntType = |
| 2009 | mlir::cast<cir::IntType>(Val: elementTypeIfVectorOrMatrix(type: dstType)); |
| 2010 | rewriter.replaceOp(op: castOp, newValues: getLLVMIntCast(rewriter, llvmSrc: llvmSrcVal, llvmDstIntTy: llvmDstType, |
| 2011 | isUnsigned: srcIntType.isUnsigned(), |
| 2012 | cirSrcWidth: srcIntType.getWidth(), |
| 2013 | cirDstIntWidth: dstIntType.getWidth())); |
| 2014 | break; |
| 2015 | } |
| 2016 | case cir::CastKind::floating: { |
| 2017 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2018 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: castOp.getType()); |
| 2019 | |
| 2020 | mlir::Type srcTy = elementTypeIfVectorOrMatrix(type: castOp.getSrc().getType()); |
| 2021 | mlir::Type dstTy = elementTypeIfVectorOrMatrix(type: castOp.getType()); |
| 2022 | |
| 2023 | if (!mlir::isa<cir::FPTypeInterface>(Val: dstTy) || |
| 2024 | !mlir::isa<cir::FPTypeInterface>(Val: srcTy)) |
| 2025 | return castOp.emitError() << "NYI cast from " << srcTy << " to " << dstTy; |
| 2026 | |
| 2027 | auto getFloatWidth = [](mlir::Type ty) -> unsigned { |
| 2028 | return mlir::cast<cir::FPTypeInterface>(Val&: ty).getWidth(); |
| 2029 | }; |
| 2030 | |
| 2031 | bool isTrunc = getFloatWidth(srcTy) > getFloatWidth(dstTy); |
| 2032 | if (cir::FenvAttr fenv = castOp.getFenvAttr()) { |
| 2033 | // fptrunc takes rounding mode + exception behavior; fpext takes only |
| 2034 | // exception behavior. |
| 2035 | return lowerToConstrainedFPIntrinsic( |
| 2036 | op: castOp, operands: llvmSrcVal, fenv, llvmResTy: llvmDstTy, rewriter, |
| 2037 | constrainedMnemonic: isTrunc ? "fptrunc" : "fpext" , /*hasRoundingMode=*/isTrunc); |
| 2038 | } |
| 2039 | if (isTrunc) |
| 2040 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPTruncOp>(op: castOp, args&: llvmDstTy, |
| 2041 | args&: llvmSrcVal); |
| 2042 | else |
| 2043 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPExtOp>(op: castOp, args&: llvmDstTy, |
| 2044 | args&: llvmSrcVal); |
| 2045 | return mlir::success(); |
| 2046 | } |
| 2047 | case cir::CastKind::int_to_ptr: { |
| 2048 | auto dstTy = mlir::cast<cir::PointerType>(Val: castOp.getType()); |
| 2049 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2050 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2051 | rewriter.replaceOpWithNewOp<mlir::LLVM::IntToPtrOp>(op: castOp, args&: llvmDstTy, |
| 2052 | args&: llvmSrcVal); |
| 2053 | return mlir::success(); |
| 2054 | } |
| 2055 | case cir::CastKind::ptr_to_int: { |
| 2056 | auto dstTy = mlir::cast<cir::IntType>(Val: castOp.getType()); |
| 2057 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2058 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2059 | rewriter.replaceOpWithNewOp<mlir::LLVM::PtrToIntOp>(op: castOp, args&: llvmDstTy, |
| 2060 | args&: llvmSrcVal); |
| 2061 | return mlir::success(); |
| 2062 | } |
| 2063 | case cir::CastKind::float_to_bool: { |
| 2064 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2065 | auto kind = mlir::LLVM::FCmpPredicate::une; |
| 2066 | |
| 2067 | // Check if float is not equal to zero. getZeroAttr yields a splat |
| 2068 | // for vector source types so this also handles element-wise |
| 2069 | // float-to-bool conversions. |
| 2070 | auto zeroFloat = mlir::LLVM::ConstantOp::create( |
| 2071 | builder&: rewriter, location: castOp.getLoc(), res: llvmSrcVal.getType(), |
| 2072 | value: rewriter.getZeroAttr(type: llvmSrcVal.getType())); |
| 2073 | |
| 2074 | // Extend comparison result to either bool (C++) or int (C). |
| 2075 | rewriter.replaceOpWithNewOp<mlir::LLVM::FCmpOp>(op: castOp, args&: kind, args&: llvmSrcVal, |
| 2076 | args&: zeroFloat); |
| 2077 | |
| 2078 | return mlir::success(); |
| 2079 | } |
| 2080 | case cir::CastKind::bool_to_int: { |
| 2081 | mlir::Type dstTy = castOp.getType(); |
| 2082 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2083 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2084 | // Compare element widths so this also handles vector bool -> int casts. |
| 2085 | auto srcElemTy = mlir::cast<mlir::IntegerType>( |
| 2086 | Val: elementTypeIfVectorOrMatrix(type: llvmSrcVal.getType())); |
| 2087 | auto dstElemTy = |
| 2088 | mlir::cast<cir::IntType>(Val: elementTypeIfVectorOrMatrix(type: dstTy)); |
| 2089 | |
| 2090 | if (srcElemTy.getWidth() == dstElemTy.getWidth()) |
| 2091 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(op: castOp, args&: llvmDstTy, |
| 2092 | args&: llvmSrcVal); |
| 2093 | else |
| 2094 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZExtOp>(op: castOp, args&: llvmDstTy, |
| 2095 | args&: llvmSrcVal); |
| 2096 | return mlir::success(); |
| 2097 | } |
| 2098 | case cir::CastKind::bool_to_float: { |
| 2099 | mlir::Type dstTy = castOp.getType(); |
| 2100 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2101 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2102 | rewriter.replaceOpWithNewOp<mlir::LLVM::UIToFPOp>(op: castOp, args&: llvmDstTy, |
| 2103 | args&: llvmSrcVal); |
| 2104 | return mlir::success(); |
| 2105 | } |
| 2106 | case cir::CastKind::int_to_float: { |
| 2107 | mlir::Type dstTy = castOp.getType(); |
| 2108 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2109 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2110 | bool isSigned = mlir::cast<cir::IntType>( |
| 2111 | Val: elementTypeIfVectorOrMatrix(type: castOp.getSrc().getType())) |
| 2112 | .isSigned(); |
| 2113 | if (cir::FenvAttr fenv = castOp.getFenvAttr()) { |
| 2114 | return lowerToConstrainedFPIntrinsic( |
| 2115 | op: castOp, operands: llvmSrcVal, fenv, llvmResTy: llvmDstTy, rewriter, |
| 2116 | constrainedMnemonic: isSigned ? "sitofp" : "uitofp" , /*hasRoundingMode=*/true); |
| 2117 | } |
| 2118 | if (isSigned) |
| 2119 | rewriter.replaceOpWithNewOp<mlir::LLVM::SIToFPOp>(op: castOp, args&: llvmDstTy, |
| 2120 | args&: llvmSrcVal); |
| 2121 | else |
| 2122 | rewriter.replaceOpWithNewOp<mlir::LLVM::UIToFPOp>(op: castOp, args&: llvmDstTy, |
| 2123 | args&: llvmSrcVal); |
| 2124 | return mlir::success(); |
| 2125 | } |
| 2126 | case cir::CastKind::float_to_int: { |
| 2127 | mlir::Type dstTy = castOp.getType(); |
| 2128 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2129 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2130 | bool isSigned = |
| 2131 | mlir::cast<cir::IntType>(Val: elementTypeIfVectorOrMatrix(type: castOp.getType())) |
| 2132 | .isSigned(); |
| 2133 | if (cir::FenvAttr fenv = castOp.getFenvAttr()) { |
| 2134 | return lowerToConstrainedFPIntrinsic( |
| 2135 | op: castOp, operands: llvmSrcVal, fenv, llvmResTy: llvmDstTy, rewriter, |
| 2136 | constrainedMnemonic: isSigned ? "fptosi" : "fptoui" , /*hasRoundingMode=*/false); |
| 2137 | } |
| 2138 | if (isSigned) |
| 2139 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPToSIOp>(op: castOp, args&: llvmDstTy, |
| 2140 | args&: llvmSrcVal); |
| 2141 | else |
| 2142 | rewriter.replaceOpWithNewOp<mlir::LLVM::FPToUIOp>(op: castOp, args&: llvmDstTy, |
| 2143 | args&: llvmSrcVal); |
| 2144 | return mlir::success(); |
| 2145 | } |
| 2146 | case cir::CastKind::bitcast: { |
| 2147 | mlir::Type dstTy = castOp.getType(); |
| 2148 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2149 | |
| 2150 | assert(!MissingFeatures::cxxABI()); |
| 2151 | assert(!MissingFeatures::dataMemberType()); |
| 2152 | |
| 2153 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2154 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(op: castOp, args&: llvmDstTy, |
| 2155 | args&: llvmSrcVal); |
| 2156 | return mlir::success(); |
| 2157 | } |
| 2158 | case cir::CastKind::ptr_to_bool: { |
| 2159 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2160 | mlir::Value zeroPtr = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: castOp.getLoc(), |
| 2161 | res: llvmSrcVal.getType()); |
| 2162 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
| 2163 | op: castOp, args: mlir::LLVM::ICmpPredicate::ne, args&: llvmSrcVal, args&: zeroPtr); |
| 2164 | break; |
| 2165 | } |
| 2166 | case cir::CastKind::address_space: { |
| 2167 | mlir::Type dstTy = castOp.getType(); |
| 2168 | mlir::Value llvmSrcVal = adaptor.getSrc(); |
| 2169 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 2170 | rewriter.replaceOpWithNewOp<mlir::LLVM::AddrSpaceCastOp>(op: castOp, args&: llvmDstTy, |
| 2171 | args&: llvmSrcVal); |
| 2172 | break; |
| 2173 | } |
| 2174 | case cir::CastKind::member_ptr_to_bool: |
| 2175 | assert(!MissingFeatures::cxxABI()); |
| 2176 | assert(!MissingFeatures::methodType()); |
| 2177 | break; |
| 2178 | default: { |
| 2179 | return castOp.emitError(message: "Unhandled cast kind: " ) |
| 2180 | << castOp.getKindAttrName(); |
| 2181 | } |
| 2182 | } |
| 2183 | |
| 2184 | return mlir::success(); |
| 2185 | } |
| 2186 | |
| 2187 | mlir::LogicalResult CIRToLLVMBuiltinIntCastOpLowering::matchAndRewrite( |
| 2188 | cir::BuiltinIntCastOp op, OpAdaptor adaptor, |
| 2189 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2190 | // Both the CIR integer and the builtin integer/index lower to LLVM integer |
| 2191 | // types, so this cast becomes an integer resize. Signedness is taken from |
| 2192 | // the CIR integer side (the builtin/index side is treated as signless). |
| 2193 | bool isUnsigned = true; |
| 2194 | if (auto cirSrc = mlir::dyn_cast<cir::IntType>(Val: op.getSrc().getType())) |
| 2195 | isUnsigned = cirSrc.isUnsigned(); |
| 2196 | else if (auto cirDst = mlir::dyn_cast<cir::IntType>(Val: op.getType())) |
| 2197 | isUnsigned = cirDst.isUnsigned(); |
| 2198 | |
| 2199 | mlir::Value llvmSrc = adaptor.getSrc(); |
| 2200 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: op.getType()); |
| 2201 | auto srcIntTy = mlir::cast<mlir::IntegerType>(Val: llvmSrc.getType()); |
| 2202 | auto dstIntTy = mlir::cast<mlir::IntegerType>(Val&: llvmDstTy); |
| 2203 | unsigned srcWidth = srcIntTy.getWidth(); |
| 2204 | unsigned dstWidth = dstIntTy.getWidth(); |
| 2205 | |
| 2206 | // Fixed-width builtin integers must match the CIR integer width. |
| 2207 | // If the converted LLVM widths differ, the non-CIR side must have been |
| 2208 | // 'index' type (target dependent width). |
| 2209 | assert((srcWidth == dstWidth || |
| 2210 | mlir::isa<mlir::IndexType>(op.getSrc().getType()) || |
| 2211 | mlir::isa<mlir::IndexType>(op.getType())) && |
| 2212 | "only index casts may change width during lowering" ); |
| 2213 | |
| 2214 | // For equal widths getLLVMIntCast returns the source unchanged, so casts |
| 2215 | // between CIR integers and fixed-width builtin integers lower to a no-op. |
| 2216 | rewriter.replaceOp(op, newValues: getLLVMIntCast(rewriter, llvmSrc, llvmDstIntTy: dstIntTy, isUnsigned, |
| 2217 | cirSrcWidth: srcWidth, cirDstIntWidth: dstWidth)); |
| 2218 | return mlir::success(); |
| 2219 | } |
| 2220 | |
| 2221 | static mlir::Value convertToIndexTy(mlir::ConversionPatternRewriter &rewriter, |
| 2222 | mlir::ModuleOp mod, mlir::Value index, |
| 2223 | mlir::Type baseTy, cir::IntType strideTy) { |
| 2224 | mlir::Operation *indexOp = index.getDefiningOp(); |
| 2225 | if (!indexOp) |
| 2226 | return index; |
| 2227 | |
| 2228 | auto indexType = mlir::cast<mlir::IntegerType>(Val: index.getType()); |
| 2229 | mlir::DataLayout llvmLayout(mod); |
| 2230 | std::optional<uint64_t> layoutWidth = llvmLayout.getTypeIndexBitwidth(t: baseTy); |
| 2231 | |
| 2232 | // If there is no change in width, don't do anything. |
| 2233 | if (!layoutWidth || *layoutWidth == indexType.getWidth()) |
| 2234 | return index; |
| 2235 | |
| 2236 | // If the index comes from a subtraction, make sure the extension happens |
| 2237 | // before it. To achieve that, look at unary minus, which already got |
| 2238 | // lowered to "sub 0, x". |
| 2239 | auto sub = dyn_cast<mlir::LLVM::SubOp>(Val: indexOp); |
| 2240 | bool rewriteSub = false; |
| 2241 | if (sub) { |
| 2242 | if (auto lhsConst = dyn_cast_if_present<mlir::LLVM::ConstantOp>( |
| 2243 | Val: sub.getLhs().getDefiningOp())) { |
| 2244 | auto lhsConstInt = mlir::dyn_cast<mlir::IntegerAttr>(Val: lhsConst.getValue()); |
| 2245 | if (lhsConstInt && lhsConstInt.getValue() == 0) { |
| 2246 | index = sub.getRhs(); |
| 2247 | rewriteSub = true; |
| 2248 | } |
| 2249 | } |
| 2250 | } |
| 2251 | |
| 2252 | auto llvmDstType = rewriter.getIntegerType(width: *layoutWidth); |
| 2253 | bool isUnsigned = strideTy && strideTy.isUnsigned(); |
| 2254 | index = getLLVMIntCast(rewriter, llvmSrc: index, llvmDstIntTy: llvmDstType, isUnsigned, |
| 2255 | cirSrcWidth: indexType.getWidth(), cirDstIntWidth: *layoutWidth); |
| 2256 | |
| 2257 | if (rewriteSub) { |
| 2258 | index = mlir::LLVM::SubOp::create( |
| 2259 | builder&: rewriter, location: index.getLoc(), |
| 2260 | lhs: mlir::LLVM::ConstantOp::create(builder&: rewriter, location: index.getLoc(), |
| 2261 | type: index.getType(), value: 0), |
| 2262 | rhs: index); |
| 2263 | } |
| 2264 | |
| 2265 | return index; |
| 2266 | } |
| 2267 | |
| 2268 | mlir::LogicalResult CIRToLLVMPtrStrideOpLowering::matchAndRewrite( |
| 2269 | cir::PtrStrideOp ptrStrideOp, OpAdaptor adaptor, |
| 2270 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2271 | |
| 2272 | const mlir::TypeConverter *tc = getTypeConverter(); |
| 2273 | const mlir::Type resultTy = tc->convertType(t: ptrStrideOp.getType()); |
| 2274 | |
| 2275 | mlir::Type elementTy = |
| 2276 | convertTypeForMemory(converter: *tc, dataLayout, type: ptrStrideOp.getElementType()); |
| 2277 | if (!elementTy) |
| 2278 | return emitTypeConversionError(op: ptrStrideOp, type: ptrStrideOp.getElementType()); |
| 2279 | |
| 2280 | // void and function types doesn't really have a layout to use in GEPs, |
| 2281 | // make it i8 instead. |
| 2282 | if (mlir::isa<mlir::LLVM::LLVMVoidType>(Val: elementTy) || |
| 2283 | mlir::isa<mlir::LLVM::LLVMFunctionType>(Val: elementTy)) |
| 2284 | elementTy = mlir::IntegerType::get(context: elementTy.getContext(), width: 8, |
| 2285 | signedness: mlir::IntegerType::Signless); |
| 2286 | // Zero-extend, sign-extend or trunc the pointer value. |
| 2287 | mlir::Value index = adaptor.getStride(); |
| 2288 | index = convertToIndexTy( |
| 2289 | rewriter, mod: ptrStrideOp->getParentOfType<mlir::ModuleOp>(), index, |
| 2290 | baseTy: adaptor.getBase().getType(), |
| 2291 | strideTy: dyn_cast<cir::IntType>(Val: ptrStrideOp.getOperand(i: 1).getType())); |
| 2292 | |
| 2293 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 2294 | op: ptrStrideOp, args: resultTy, args&: elementTy, args: adaptor.getBase(), args&: index); |
| 2295 | return mlir::success(); |
| 2296 | } |
| 2297 | |
| 2298 | mlir::LogicalResult CIRToLLVMGetElementOpLowering::matchAndRewrite( |
| 2299 | cir::GetElementOp op, OpAdaptor adaptor, |
| 2300 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2301 | if (auto arrayTy = |
| 2302 | mlir::dyn_cast<cir::ArrayType>(Val: op.getBaseType().getPointee())) { |
| 2303 | const mlir::TypeConverter *converter = getTypeConverter(); |
| 2304 | const mlir::Type llArrayTy = converter->convertType(t: arrayTy); |
| 2305 | if (!llArrayTy) |
| 2306 | return emitTypeConversionError(op, type: arrayTy); |
| 2307 | const mlir::Type llResultTy = converter->convertType(t: op.getType()); |
| 2308 | mlir::Type elementTy = |
| 2309 | convertTypeForMemory(converter: *converter, dataLayout, type: op.getElementType()); |
| 2310 | |
| 2311 | // void and function types don't really have a layout to use in GEPs, |
| 2312 | // make it i8 instead. |
| 2313 | if (mlir::isa<mlir::LLVM::LLVMVoidType>(Val: elementTy) || |
| 2314 | mlir::isa<mlir::LLVM::LLVMFunctionType>(Val: elementTy)) |
| 2315 | elementTy = rewriter.getIntegerType(width: 8); |
| 2316 | |
| 2317 | mlir::Value index = adaptor.getIndex(); |
| 2318 | index = |
| 2319 | convertToIndexTy(rewriter, mod: op->getParentOfType<mlir::ModuleOp>(), index, |
| 2320 | baseTy: adaptor.getBase().getType(), |
| 2321 | strideTy: dyn_cast<cir::IntType>(Val: op.getOperand(i: 1).getType())); |
| 2322 | |
| 2323 | // Since the base address is a pointer to an array, the first index is |
| 2324 | // always zero, and the second one selects the element. The zero takes the |
| 2325 | // element index's type, because a raw constant GEP index is always |
| 2326 | // translated as i32. |
| 2327 | mlir::Value zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 2328 | type: index.getType(), value: 0); |
| 2329 | std::array<mlir::LLVM::GEPArg, 2> offset = {zero, index}; |
| 2330 | mlir::LLVM::GEPNoWrapFlags flags = mlir::LLVM::GEPNoWrapFlags::none; |
| 2331 | if (op.getInBounds()) |
| 2332 | flags |= mlir::LLVM::GEPNoWrapFlags::inbounds; |
| 2333 | if (op.getNoUnsignedWrap()) |
| 2334 | flags |= mlir::LLVM::GEPNoWrapFlags::nuw; |
| 2335 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 2336 | op, args: llResultTy, args: llArrayTy, args: adaptor.getBase(), args&: offset, args&: flags); |
| 2337 | return mlir::success(); |
| 2338 | } |
| 2339 | |
| 2340 | op.emitError() << "NYI: GetElementOp lowering to LLVM for non-array" ; |
| 2341 | return mlir::failure(); |
| 2342 | } |
| 2343 | |
| 2344 | mlir::LogicalResult CIRToLLVMBaseClassAddrOpLowering::matchAndRewrite( |
| 2345 | cir::BaseClassAddrOp baseClassOp, OpAdaptor adaptor, |
| 2346 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2347 | const mlir::Type resultType = |
| 2348 | getTypeConverter()->convertType(t: baseClassOp.getType()); |
| 2349 | if (!resultType) |
| 2350 | return emitTypeConversionError(op: baseClassOp, type: baseClassOp.getType()); |
| 2351 | mlir::Value derivedAddr = adaptor.getSrcAddr(); |
| 2352 | llvm::SmallVector<mlir::LLVM::GEPArg, 1> offset = { |
| 2353 | static_cast<int32_t>(adaptor.getOffset().getZExtValue())}; |
| 2354 | mlir::Type byteType = mlir::IntegerType::get(context: resultType.getContext(), width: 8, |
| 2355 | signedness: mlir::IntegerType::Signless); |
| 2356 | if (adaptor.getOffset().getZExtValue() == 0) { |
| 2357 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(op: baseClassOp, args: resultType, |
| 2358 | args: adaptor.getSrcAddr()); |
| 2359 | return mlir::success(); |
| 2360 | } |
| 2361 | |
| 2362 | if (baseClassOp.getAssumeNotNull()) { |
| 2363 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 2364 | op: baseClassOp, args: resultType, args&: byteType, args&: derivedAddr, args&: offset); |
| 2365 | } else { |
| 2366 | auto loc = baseClassOp.getLoc(); |
| 2367 | mlir::Value isNull = mlir::LLVM::ICmpOp::create( |
| 2368 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::eq, lhs: derivedAddr, |
| 2369 | rhs: mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: derivedAddr.getType())); |
| 2370 | mlir::Value adjusted = mlir::LLVM::GEPOp::create( |
| 2371 | builder&: rewriter, location: loc, resultType, elementType: byteType, basePtr: derivedAddr, indices: offset); |
| 2372 | rewriter.replaceOpWithNewOp<mlir::LLVM::SelectOp>(op: baseClassOp, args&: isNull, |
| 2373 | args&: derivedAddr, args&: adjusted); |
| 2374 | } |
| 2375 | return mlir::success(); |
| 2376 | } |
| 2377 | |
| 2378 | mlir::LogicalResult CIRToLLVMDerivedClassAddrOpLowering::matchAndRewrite( |
| 2379 | cir::DerivedClassAddrOp derivedClassOp, OpAdaptor adaptor, |
| 2380 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2381 | const mlir::Type resultType = |
| 2382 | getTypeConverter()->convertType(t: derivedClassOp.getType()); |
| 2383 | if (!resultType) |
| 2384 | return emitTypeConversionError(op: derivedClassOp, type: derivedClassOp.getType()); |
| 2385 | mlir::Value baseAddr = adaptor.getSrcAddr(); |
| 2386 | // The offset is set in the operation as an unsigned value, but it must be |
| 2387 | // applied as a negative offset. |
| 2388 | int64_t offsetVal = -(adaptor.getOffset().getZExtValue()); |
| 2389 | if (offsetVal == 0) { |
| 2390 | // If the offset is zero, we can just return the source address. |
| 2391 | rewriter.replaceOp(op: derivedClassOp, newValues: baseAddr); |
| 2392 | return mlir::success(); |
| 2393 | } |
| 2394 | llvm::SmallVector<mlir::LLVM::GEPArg, 1> offset = { |
| 2395 | static_cast<int32_t>(offsetVal)}; |
| 2396 | mlir::Type byteType = mlir::IntegerType::get(context: resultType.getContext(), width: 8, |
| 2397 | signedness: mlir::IntegerType::Signless); |
| 2398 | if (derivedClassOp.getAssumeNotNull()) { |
| 2399 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 2400 | op: derivedClassOp, args: resultType, args&: byteType, args&: baseAddr, args&: offset, |
| 2401 | args: mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 2402 | } else { |
| 2403 | mlir::Location loc = derivedClassOp.getLoc(); |
| 2404 | mlir::Value isNull = mlir::LLVM::ICmpOp::create( |
| 2405 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::eq, lhs: baseAddr, |
| 2406 | rhs: mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: baseAddr.getType())); |
| 2407 | mlir::Value adjusted = |
| 2408 | mlir::LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType, elementType: byteType, basePtr: baseAddr, |
| 2409 | indices: offset, noWrapFlags: mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 2410 | rewriter.replaceOpWithNewOp<mlir::LLVM::SelectOp>(op: derivedClassOp, args&: isNull, |
| 2411 | args&: baseAddr, args&: adjusted); |
| 2412 | } |
| 2413 | return mlir::success(); |
| 2414 | } |
| 2415 | |
| 2416 | mlir::LogicalResult CIRToLLVMFMaxNumOpLowering::matchAndRewrite( |
| 2417 | cir::FMaxNumOp op, OpAdaptor adaptor, |
| 2418 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2419 | mlir::Type resTy = typeConverter->convertType(t: op.getType()); |
| 2420 | if (cir::FenvAttr fenv = op.getFenvAttr()) |
| 2421 | return lowerToConstrainedFPIntrinsic( |
| 2422 | op, operands: adaptor.getOperands(), fenv, llvmResTy: resTy, rewriter, constrainedMnemonic: "maxnum" , |
| 2423 | /*hasRoundingMode=*/false, fastmathFlags: mlir::LLVM::FastmathFlags::nsz); |
| 2424 | rewriter.replaceOpWithNewOp<mlir::LLVM::MaxNumOp>( |
| 2425 | op, args&: resTy, args: adaptor.getLhs(), args: adaptor.getRhs(), |
| 2426 | args: mlir::LLVM::FastmathFlags::nsz); |
| 2427 | return mlir::success(); |
| 2428 | } |
| 2429 | |
| 2430 | mlir::LogicalResult CIRToLLVMFMinNumOpLowering::matchAndRewrite( |
| 2431 | cir::FMinNumOp op, OpAdaptor adaptor, |
| 2432 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2433 | mlir::Type resTy = typeConverter->convertType(t: op.getType()); |
| 2434 | if (cir::FenvAttr fenv = op.getFenvAttr()) |
| 2435 | return lowerToConstrainedFPIntrinsic( |
| 2436 | op, operands: adaptor.getOperands(), fenv, llvmResTy: resTy, rewriter, constrainedMnemonic: "minnum" , |
| 2437 | /*hasRoundingMode=*/false, fastmathFlags: mlir::LLVM::FastmathFlags::nsz); |
| 2438 | rewriter.replaceOpWithNewOp<mlir::LLVM::MinNumOp>( |
| 2439 | op, args&: resTy, args: adaptor.getLhs(), args: adaptor.getRhs(), |
| 2440 | args: mlir::LLVM::FastmathFlags::nsz); |
| 2441 | return mlir::success(); |
| 2442 | } |
| 2443 | |
| 2444 | mlir::LogicalResult CIRToLLVMAllocaOpLowering::matchAndRewrite( |
| 2445 | cir::AllocaOp op, OpAdaptor adaptor, |
| 2446 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2447 | mlir::ModuleOp module = op->getParentOfType<mlir::ModuleOp>(); |
| 2448 | mlir::Attribute langOpts = |
| 2449 | module->getAttr(name: cir::CIRDialect::getLoweringOptionsAttrName()); |
| 2450 | auto loweringLangOps = |
| 2451 | dyn_cast_if_present<cir::LoweringOptionsAttr>(Val&: langOpts); |
| 2452 | |
| 2453 | bool isForHLSL = loweringLangOps ? loweringLangOps.getHlsl() : false; |
| 2454 | if (isForHLSL && mlir::isa<cir::MatrixType>(Val: op.getAllocaType()) && |
| 2455 | mlir::isa<cir::BoolType>( |
| 2456 | Val: cast<cir::MatrixType>(Val: op.getAllocaType()).getElementType())) { |
| 2457 | return op.emitError() |
| 2458 | << "NYI: Builtin matrix type with bool element type for HLSL" ; |
| 2459 | } |
| 2460 | |
| 2461 | mlir::Value size = |
| 2462 | op.isDynamic() |
| 2463 | ? adaptor.getDynAllocSize() |
| 2464 | : mlir::LLVM::ConstantOp::create( |
| 2465 | builder&: rewriter, location: op.getLoc(), |
| 2466 | type: typeConverter->convertType(t: rewriter.getI32Type()), value: 1); |
| 2467 | mlir::Type elementTy = convertTypeForMemory(converter: *getTypeConverter(), dataLayout, |
| 2468 | type: op.getAllocaType(), isForHLSL); |
| 2469 | if (!elementTy) |
| 2470 | return op.emitError() |
| 2471 | << "NYI: lowering alloca of a type with no memory representation" ; |
| 2472 | mlir::Type resultTy = convertTypeForMemory(converter: *getTypeConverter(), dataLayout, |
| 2473 | type: op.getType(), isForHLSL); |
| 2474 | if (!resultTy) |
| 2475 | return emitTypeConversionError(op, type: op.getType()); |
| 2476 | |
| 2477 | assert(!cir::MissingFeatures::addressSpace()); |
| 2478 | |
| 2479 | rewriter.replaceOpWithNewOp<mlir::LLVM::AllocaOp>(op, args&: resultTy, args&: elementTy, |
| 2480 | args&: size, args: op.getAlignment()); |
| 2481 | |
| 2482 | return mlir::success(); |
| 2483 | } |
| 2484 | |
| 2485 | mlir::LogicalResult CIRToLLVMRotateOpLowering::matchAndRewrite( |
| 2486 | cir::RotateOp op, OpAdaptor adaptor, |
| 2487 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2488 | // Note that LLVM intrinsic calls to @llvm.fsh{r,l}.i* have the same type as |
| 2489 | // the operand. |
| 2490 | mlir::Value input = adaptor.getInput(); |
| 2491 | if (op.isRotateLeft()) |
| 2492 | rewriter.replaceOpWithNewOp<mlir::LLVM::FshlOp>(op, args&: input, args&: input, |
| 2493 | args: adaptor.getAmount()); |
| 2494 | else |
| 2495 | rewriter.replaceOpWithNewOp<mlir::LLVM::FshrOp>(op, args&: input, args&: input, |
| 2496 | args: adaptor.getAmount()); |
| 2497 | return mlir::LogicalResult::success(); |
| 2498 | } |
| 2499 | |
| 2500 | /// The `llvm.byval`, `llvm.sret`, and `llvm.byref` argument attributes carry |
| 2501 | /// the pointee type as a TypeAttr. After the CallConvLowering pass that type |
| 2502 | /// is still a CIR type; remap it to the lowered LLVM type so translation to |
| 2503 | /// LLVM IR does not encounter a CIR type in an attribute. The pointee must |
| 2504 | /// name the in-memory storage type, not the value type, since it sizes the |
| 2505 | /// byval copy and the sret slot: a _BitInt narrower than its padded storage |
| 2506 | /// integer would otherwise understate both. Returns std::nullopt when a |
| 2507 | /// pointee has no memory representation, leaving the caller to emit the NYI |
| 2508 | /// error, and the input unchanged when there is nothing to convert. |
| 2509 | static std::optional<mlir::ArrayAttr> convertTypedArgAttrs( |
| 2510 | mlir::ArrayAttr argAttrs, const mlir::TypeConverter &converter, |
| 2511 | const mlir::DataLayout &dataLayout, mlir::MLIRContext *ctx) { |
| 2512 | if (!argAttrs) |
| 2513 | return argAttrs; |
| 2514 | bool changed = false; |
| 2515 | SmallVector<mlir::Attribute> loweredArgAttrs; |
| 2516 | loweredArgAttrs.reserve(N: argAttrs.size()); |
| 2517 | for (mlir::Attribute a : argAttrs) { |
| 2518 | auto dict = cast<mlir::DictionaryAttr>(Val&: a); |
| 2519 | SmallVector<mlir::NamedAttribute> entries(dict.begin(), dict.end()); |
| 2520 | for (mlir::NamedAttribute &entry : entries) { |
| 2521 | StringRef name = entry.getName().strref(); |
| 2522 | if (name != mlir::LLVM::LLVMDialect::getByValAttrName() && |
| 2523 | name != mlir::LLVM::LLVMDialect::getStructRetAttrName() && |
| 2524 | name != mlir::LLVM::LLVMDialect::getByRefAttrName()) |
| 2525 | continue; |
| 2526 | auto typeAttr = dyn_cast<mlir::TypeAttr>(Val: entry.getValue()); |
| 2527 | if (!typeAttr) |
| 2528 | continue; |
| 2529 | mlir::Type lowered = |
| 2530 | convertTypeForMemory(converter, dataLayout, type: typeAttr.getValue()); |
| 2531 | if (!lowered) |
| 2532 | return std::nullopt; |
| 2533 | if (lowered != typeAttr.getValue()) { |
| 2534 | entry.setValue(mlir::TypeAttr::get(type: lowered)); |
| 2535 | changed = true; |
| 2536 | } |
| 2537 | } |
| 2538 | loweredArgAttrs.push_back(Elt: mlir::DictionaryAttr::get(context: ctx, value: entries)); |
| 2539 | } |
| 2540 | return changed ? mlir::ArrayAttr::get(context: ctx, value: loweredArgAttrs) : argAttrs; |
| 2541 | } |
| 2542 | |
| 2543 | struct LoweredCallAttributes { |
| 2544 | mlir::ArrayAttr argAttrs; |
| 2545 | mlir::ArrayAttr resAttrs; |
| 2546 | SmallVector<mlir::NamedAttribute> discardableAttrs; |
| 2547 | }; |
| 2548 | |
| 2549 | static mlir::LogicalResult lowerCallAttributes( |
| 2550 | cir::CIRCallOpInterface op, const mlir::TypeConverter &converter, |
| 2551 | const mlir::DataLayout &dataLayout, LoweredCallAttributes &result) { |
| 2552 | if (auto argAttrs = mlir::dyn_cast_if_present<mlir::ArrayAttr>( |
| 2553 | Val: op->getInherentAttr(name: CIRDialect::getArgAttrsAttrName()) |
| 2554 | .value_or(u: mlir::Attribute{}))) { |
| 2555 | std::optional<mlir::ArrayAttr> lowered = |
| 2556 | convertTypedArgAttrs(argAttrs, converter, dataLayout, ctx: op->getContext()); |
| 2557 | if (!lowered) |
| 2558 | return op->emitError() << "NYI: lowering a byval/sret/byref call " |
| 2559 | "argument whose pointee type has no " |
| 2560 | "memory representation" ; |
| 2561 | result.argAttrs = *lowered; |
| 2562 | } |
| 2563 | result.resAttrs = mlir::dyn_cast_if_present<mlir::ArrayAttr>( |
| 2564 | Val: op->getInherentAttr(name: CIRDialect::getResAttrsAttrName()) |
| 2565 | .value_or(u: mlir::Attribute{})); |
| 2566 | |
| 2567 | for (mlir::NamedAttribute attr : op->getDiscardableAttrs()) { |
| 2568 | if (attr.getName() == CIRDialect::getNoUnwindAttrName() || |
| 2569 | attr.getName() == CIRDialect::getNoReturnAttrName()) |
| 2570 | continue; |
| 2571 | assert(!cir::MissingFeatures::opFuncExtraAttrs()); |
| 2572 | result.discardableAttrs.push_back(Elt: attr); |
| 2573 | } |
| 2574 | return mlir::success(); |
| 2575 | } |
| 2576 | |
| 2577 | template <typename CallLikeOp> |
| 2578 | static void setLoweredCallAttributes(CallLikeOp op, |
| 2579 | const LoweredCallAttributes &attributes) { |
| 2580 | if (attributes.argAttrs) |
| 2581 | op.setArgAttrsAttr(attributes.argAttrs); |
| 2582 | if (attributes.resAttrs) |
| 2583 | op.setResAttrsAttr(attributes.resAttrs); |
| 2584 | for (mlir::NamedAttribute attr : attributes.discardableAttrs) { |
| 2585 | if (op->getInherentAttr(attr.getName()).has_value()) |
| 2586 | op->setInherentAttr(attr.getName(), attr.getValue()); |
| 2587 | else |
| 2588 | op->setDiscardableAttr(attr.getName(), attr.getValue()); |
| 2589 | } |
| 2590 | } |
| 2591 | |
| 2592 | static mlir::LogicalResult rewriteCallOrInvoke( |
| 2593 | mlir::Operation *op, mlir::ValueRange callOperands, |
| 2594 | mlir::ConversionPatternRewriter &rewriter, |
| 2595 | const mlir::TypeConverter *converter, const mlir::DataLayout &dataLayout, |
| 2596 | mlir::SymbolTableCollection &symbolTables, |
| 2597 | mlir::FlatSymbolRefAttr calleeAttr, mlir::Block *continueBlock = nullptr, |
| 2598 | mlir::Block *landingPadBlock = nullptr) { |
| 2599 | llvm::SmallVector<mlir::Type, 8> llvmResults; |
| 2600 | mlir::ValueTypeRange<mlir::ResultRange> cirResults = op->getResultTypes(); |
| 2601 | auto call = cast<cir::CIRCallOpInterface>(Val: op); |
| 2602 | |
| 2603 | if (converter->convertTypes(types: cirResults, results&: llvmResults).failed()) |
| 2604 | return mlir::failure(); |
| 2605 | |
| 2606 | mlir::LLVM::MemoryEffectsAttr memoryEffects; |
| 2607 | bool noUnwind = false; |
| 2608 | bool willReturn = false; |
| 2609 | bool noReturn = false; |
| 2610 | convertCallEffects(callOp: op, isNothrow: call.getNothrow(), effects: call.getMemoryEffectsAttr(), |
| 2611 | memoryEffect&: memoryEffects, noUnwind, willReturn, noReturn); |
| 2612 | |
| 2613 | LoweredCallAttributes attributes; |
| 2614 | if (mlir::failed( |
| 2615 | Result: lowerCallAttributes(op: call, converter: *converter, dataLayout, result&: attributes))) |
| 2616 | return mlir::failure(); |
| 2617 | |
| 2618 | mlir::LLVM::LLVMFunctionType llvmFnTy; |
| 2619 | |
| 2620 | // Temporary to handle the case where we need to prepend an operand if the |
| 2621 | // callee is an alias. |
| 2622 | SmallVector<mlir::Value> adjustedCallOperands; |
| 2623 | |
| 2624 | if (calleeAttr) { // direct call |
| 2625 | mlir::Operation *callee = |
| 2626 | symbolTables.lookupNearestSymbolFrom(from: op, symbol: calleeAttr); |
| 2627 | if (auto fn = mlir::dyn_cast<mlir::FunctionOpInterface>(Val: callee)) { |
| 2628 | llvmFnTy = converter->convertType<mlir::LLVM::LLVMFunctionType>( |
| 2629 | t: fn.getFunctionType()); |
| 2630 | assert(llvmFnTy && "Failed to convert function type" ); |
| 2631 | } else if (auto alias = mlir::cast<mlir::LLVM::AliasOp>(Val: callee)) { |
| 2632 | // If the callee was an alias. In that case, |
| 2633 | // we need to prepend the address of the alias to the operands. The |
| 2634 | // way aliases work in the LLVM dialect is a little counter-intuitive. |
| 2635 | // The AliasOp itself is a pseudo-function that returns the address of |
| 2636 | // the global value being aliased, but when we generate the call we |
| 2637 | // need to insert an operation that gets the address of the AliasOp. |
| 2638 | // This all gets sorted out when the LLVM dialect is lowered to LLVM IR. |
| 2639 | auto symAttr = mlir::cast<mlir::FlatSymbolRefAttr>(Val&: calleeAttr); |
| 2640 | auto addrOfAlias = |
| 2641 | mlir::LLVM::AddressOfOp::create( |
| 2642 | builder&: rewriter, location: op->getLoc(), |
| 2643 | res: mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()), global_name: symAttr) |
| 2644 | .getResult(); |
| 2645 | adjustedCallOperands.push_back(Elt: addrOfAlias); |
| 2646 | |
| 2647 | // Now add the regular operands and assign this to the range value. |
| 2648 | llvm::append_range(C&: adjustedCallOperands, R&: callOperands); |
| 2649 | callOperands = adjustedCallOperands; |
| 2650 | |
| 2651 | // Clear the callee attribute because we're calling an alias. |
| 2652 | calleeAttr = {}; |
| 2653 | llvmFnTy = mlir::cast<mlir::LLVM::LLVMFunctionType>(Val: alias.getType()); |
| 2654 | } else { |
| 2655 | // Was this an ifunc? |
| 2656 | return op->emitError(message: "Unexpected callee type!" ); |
| 2657 | } |
| 2658 | } else { // indirect call |
| 2659 | assert(!op->getOperands().empty() && |
| 2660 | "operands list must no be empty for the indirect call" ); |
| 2661 | auto calleeTy = op->getOperands().front().getType(); |
| 2662 | auto calleePtrTy = cast<cir::PointerType>(Val&: calleeTy); |
| 2663 | auto calleeFuncTy = cast<cir::FuncType>(Val: calleePtrTy.getPointee()); |
| 2664 | llvm::append_range(C&: adjustedCallOperands, R&: callOperands); |
| 2665 | llvmFnTy = cast<mlir::LLVM::LLVMFunctionType>( |
| 2666 | Val: converter->convertType(t: calleeFuncTy)); |
| 2667 | } |
| 2668 | |
| 2669 | mlir::LLVM::CConv cconv = convertCallingConv(callingConv: call.getCallingConv()); |
| 2670 | |
| 2671 | // The replace below erases the root op, so read what the CallOp branch |
| 2672 | // needs first. |
| 2673 | const bool mustTail = op->hasAttr(name: CIRDialect::getMustTailAttrName()); |
| 2674 | const std::optional<cir::InlineKind> inlineKind = call.getInlineKind(); |
| 2675 | if (landingPadBlock) { |
| 2676 | assert(!cir::MissingFeatures::opCallInvokeAttrs()); |
| 2677 | auto newOp = rewriter.replaceOpWithNewOp<mlir::LLVM::InvokeOp>( |
| 2678 | op, args&: llvmFnTy, args&: calleeAttr, args&: callOperands, args&: continueBlock, |
| 2679 | args: mlir::ValueRange{}, args&: landingPadBlock, args: mlir::ValueRange{}); |
| 2680 | newOp.setCConv(cconv); |
| 2681 | setLoweredCallAttributes(op: newOp, attributes); |
| 2682 | } else { |
| 2683 | auto newOp = rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( |
| 2684 | op, args&: llvmFnTy, args&: calleeAttr, args&: callOperands); |
| 2685 | newOp.setCConv(cconv); |
| 2686 | setLoweredCallAttributes(op: newOp, attributes); |
| 2687 | if (memoryEffects) |
| 2688 | newOp.setMemoryEffectsAttr(memoryEffects); |
| 2689 | newOp.setNoUnwind(noUnwind); |
| 2690 | newOp.setWillReturn(willReturn); |
| 2691 | newOp.setNoreturn(noReturn); |
| 2692 | if (mustTail) |
| 2693 | newOp.setTailCallKind(mlir::LLVM::TailCallKind::MustTail); |
| 2694 | |
| 2695 | if (inlineKind) { |
| 2696 | newOp.setNoInline(*inlineKind == cir::InlineKind::NoInline); |
| 2697 | newOp.setInlineHint(*inlineKind == cir::InlineKind::InlineHint); |
| 2698 | newOp.setAlwaysInline(*inlineKind == cir::InlineKind::AlwaysInline); |
| 2699 | } |
| 2700 | } |
| 2701 | |
| 2702 | return mlir::success(); |
| 2703 | } |
| 2704 | |
| 2705 | mlir::LogicalResult CIRToLLVMCallOpLowering::matchAndRewrite( |
| 2706 | cir::CallOp op, OpAdaptor adaptor, |
| 2707 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2708 | return rewriteCallOrInvoke(op: op.getOperation(), callOperands: adaptor.getOperands(), rewriter, |
| 2709 | converter: getTypeConverter(), dataLayout, symbolTables, |
| 2710 | calleeAttr: op.getCalleeAttr()); |
| 2711 | } |
| 2712 | |
| 2713 | mlir::LogicalResult CIRToLLVMTryCallOpLowering::matchAndRewrite( |
| 2714 | cir::TryCallOp op, OpAdaptor adaptor, |
| 2715 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2716 | return rewriteCallOrInvoke(op: op.getOperation(), callOperands: adaptor.getOperands(), rewriter, |
| 2717 | converter: getTypeConverter(), dataLayout, symbolTables, |
| 2718 | calleeAttr: op.getCalleeAttr(), continueBlock: op.getNormalDest(), |
| 2719 | landingPadBlock: op.getUnwindDest()); |
| 2720 | } |
| 2721 | |
| 2722 | mlir::LogicalResult CIRToLLVMReturnAddrOpLowering::matchAndRewrite( |
| 2723 | cir::ReturnAddrOp op, OpAdaptor adaptor, |
| 2724 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2725 | const mlir::Type llvmPtrTy = getTypeConverter()->convertType(t: op.getType()); |
| 2726 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm.returnaddress" , |
| 2727 | resultTy: llvmPtrTy, operands: adaptor.getOperands()); |
| 2728 | return mlir::success(); |
| 2729 | } |
| 2730 | |
| 2731 | mlir::LogicalResult CIRToLLVMFrameAddrOpLowering::matchAndRewrite( |
| 2732 | cir::FrameAddrOp op, OpAdaptor adaptor, |
| 2733 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2734 | const mlir::Type llvmPtrTy = getTypeConverter()->convertType(t: op.getType()); |
| 2735 | if (!llvmPtrTy) |
| 2736 | return emitTypeConversionError(op, type: op.getType()); |
| 2737 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm.frameaddress" , resultTy: llvmPtrTy, |
| 2738 | operands: adaptor.getOperands()); |
| 2739 | return mlir::success(); |
| 2740 | } |
| 2741 | |
| 2742 | mlir::LogicalResult CIRToLLVMClearCacheOpLowering::matchAndRewrite( |
| 2743 | cir::ClearCacheOp op, OpAdaptor adaptor, |
| 2744 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2745 | mlir::Value begin = adaptor.getBegin(); |
| 2746 | mlir::Value end = adaptor.getEnd(); |
| 2747 | auto intrinNameAttr = |
| 2748 | mlir::StringAttr::get(context: op.getContext(), bytes: "llvm.clear_cache" ); |
| 2749 | rewriter.replaceOpWithNewOp<mlir::LLVM::CallIntrinsicOp>( |
| 2750 | op, args: mlir::Type{}, args&: intrinNameAttr, args: mlir::ValueRange{begin, end}); |
| 2751 | |
| 2752 | return mlir::success(); |
| 2753 | } |
| 2754 | |
| 2755 | mlir::LogicalResult CIRToLLVMAddrOfReturnAddrOpLowering::matchAndRewrite( |
| 2756 | cir::AddrOfReturnAddrOp op, OpAdaptor adaptor, |
| 2757 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2758 | const mlir::Type llvmPtrTy = getTypeConverter()->convertType(t: op.getType()); |
| 2759 | if (!llvmPtrTy) |
| 2760 | return emitTypeConversionError(op, type: op.getType()); |
| 2761 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm.addressofreturnaddress" , |
| 2762 | resultTy: llvmPtrTy, operands: adaptor.getOperands()); |
| 2763 | return mlir::success(); |
| 2764 | } |
| 2765 | |
| 2766 | mlir::LogicalResult CIRToLLVMAMDGPUFeaturePredicateOpLowering::matchAndRewrite( |
| 2767 | cir::AMDGPUFeaturePredicateOp op, OpAdaptor adaptor, |
| 2768 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2769 | mlir::Location loc = op.getLoc(); |
| 2770 | mlir::MLIRContext *ctx = rewriter.getContext(); |
| 2771 | mlir::Value specId = |
| 2772 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI32Type(), value: -1); |
| 2773 | mlir::Value defaultValue = |
| 2774 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI1Type(), value: 0); |
| 2775 | auto predicate = mlir::LLVM::MDNodeAttr::get( |
| 2776 | context: ctx, operands: mlir::LLVM::MDStringAttr::get(context: ctx, value: op.getPredicateAttr())); |
| 2777 | mlir::Value md = |
| 2778 | mlir::LLVM::MetadataAsValueOp::create(builder&: rewriter, location: loc, metadata: predicate); |
| 2779 | replaceOpWithCallLLVMIntrinsicOp( |
| 2780 | rewriter, op, intrinsicName: "llvm.spv.named.boolean.spec.constant" , |
| 2781 | resultTy: getTypeConverter()->convertType(t: op.getType()), |
| 2782 | operands: mlir::ValueRange{specId, defaultValue, md}); |
| 2783 | return mlir::success(); |
| 2784 | } |
| 2785 | |
| 2786 | mlir::LogicalResult CIRToLLVMLoadOpLowering::matchAndRewrite( |
| 2787 | cir::LoadOp op, OpAdaptor adaptor, |
| 2788 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2789 | const mlir::Type llvmTy = |
| 2790 | convertTypeForLoadStore(converter: *getTypeConverter(), dataLayout, type: op.getType()); |
| 2791 | if (!llvmTy) |
| 2792 | return emitTypeConversionError(op, type: op.getType()); |
| 2793 | mlir::LLVM::AtomicOrdering ordering = getLLVMMemOrder(memorder: op.getMemOrder()); |
| 2794 | std::optional<size_t> opAlign = op.getAlignment(); |
| 2795 | unsigned alignment = (unsigned)opAlign.value_or( |
| 2796 | u: getMemoryFallbackAlignment(op, cirType: op.getType(), llvmMemType: llvmTy, dataLayout)); |
| 2797 | |
| 2798 | assert(!cir::MissingFeatures::lowerModeOptLevel()); |
| 2799 | |
| 2800 | std::optional<llvm::StringRef> llvmSyncScope = getLLVMSyncScope( |
| 2801 | syncScope: op.getSyncScope(), moduleOp: op->getParentOfType<mlir::ModuleOp>()); |
| 2802 | |
| 2803 | mlir::LLVM::LoadOp newLoad = mlir::LLVM::LoadOp::create( |
| 2804 | builder&: rewriter, location: op->getLoc(), type: llvmTy, addr: adaptor.getAddr(), alignment, |
| 2805 | isVolatile: op.getIsVolatile(), /*isNonTemporal=*/op.getIsNontemporal(), |
| 2806 | /*isInvariant=*/op.getInvariant(), /*isInvariantGroup=*/false, |
| 2807 | /*isNoundef=*/false, ordering, syncscope: llvmSyncScope.value_or(u: std::string())); |
| 2808 | if (mlir::Attribute domain = op->getAttr(name: "cir.riscv_nontemporal_domain" )) |
| 2809 | newLoad->setAttr(name: "cir.riscv_nontemporal_domain" , value: domain); |
| 2810 | |
| 2811 | // Convert adapted result to its original type if needed. |
| 2812 | mlir::Value result = emitFromMemory(rewriter, converter: *getTypeConverter(), dataLayout, |
| 2813 | op, value: newLoad.getResult()); |
| 2814 | rewriter.replaceOp(op, newValues: result); |
| 2815 | assert(!cir::MissingFeatures::opLoadStoreTbaa()); |
| 2816 | return mlir::LogicalResult::success(); |
| 2817 | } |
| 2818 | |
| 2819 | mlir::LogicalResult |
| 2820 | cir::direct::CIRToLLVMVecMaskedLoadOpLowering::matchAndRewrite( |
| 2821 | cir::VecMaskedLoadOp op, OpAdaptor adaptor, |
| 2822 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2823 | const mlir::Type llvmResTy = |
| 2824 | convertTypeForLoadStore(converter: *getTypeConverter(), dataLayout, type: op.getType()); |
| 2825 | if (!llvmResTy) |
| 2826 | return op.emitError() |
| 2827 | << "NYI: lowering masked load of a type with no memory " |
| 2828 | "representation" ; |
| 2829 | |
| 2830 | std::optional<size_t> opAlign = op.getAlignment(); |
| 2831 | unsigned alignment = |
| 2832 | (unsigned)opAlign.value_or(u: dataLayout.getTypeABIAlignment(t: llvmResTy)); |
| 2833 | |
| 2834 | auto newLoad = mlir::LLVM::MaskedLoadOp::create( |
| 2835 | builder&: rewriter, location: op.getLoc(), res: llvmResTy, data: adaptor.getAddr(), mask: adaptor.getMask(), |
| 2836 | pass_thru: adaptor.getPassThru(), alignment); |
| 2837 | |
| 2838 | rewriter.replaceOp(op, newValues: newLoad.getResult()); |
| 2839 | return mlir::success(); |
| 2840 | } |
| 2841 | |
| 2842 | mlir::LogicalResult CIRToLLVMStoreOpLowering::matchAndRewrite( |
| 2843 | cir::StoreOp op, OpAdaptor adaptor, |
| 2844 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2845 | mlir::LLVM::AtomicOrdering memorder = getLLVMMemOrder(memorder: op.getMemOrder()); |
| 2846 | mlir::Type valueType = op.getValue().getType(); |
| 2847 | const mlir::Type llvmTy = |
| 2848 | convertTypeForLoadStore(converter: *getTypeConverter(), dataLayout, type: valueType); |
| 2849 | if (!llvmTy) |
| 2850 | return op.emitError() |
| 2851 | << "NYI: lowering store of a type with no memory representation" ; |
| 2852 | std::optional<size_t> opAlign = op.getAlignment(); |
| 2853 | unsigned alignment = (unsigned)opAlign.value_or( |
| 2854 | u: getMemoryFallbackAlignment(op, cirType: valueType, llvmMemType: llvmTy, dataLayout)); |
| 2855 | |
| 2856 | assert(!cir::MissingFeatures::lowerModeOptLevel()); |
| 2857 | |
| 2858 | // Convert adapted value to its memory type if needed. |
| 2859 | mlir::Value value = emitToMemory(rewriter, dataLayout, |
| 2860 | origType: op.getValue().getType(), value: adaptor.getValue()); |
| 2861 | assert(!cir::MissingFeatures::opLoadStoreTbaa()); |
| 2862 | |
| 2863 | std::optional<llvm::StringRef> llvmSyncScope = getLLVMSyncScope( |
| 2864 | syncScope: op.getSyncScope(), moduleOp: op->getParentOfType<mlir::ModuleOp>()); |
| 2865 | |
| 2866 | mlir::LLVM::StoreOp storeOp = mlir::LLVM::StoreOp::create( |
| 2867 | builder&: rewriter, location: op->getLoc(), value, addr: adaptor.getAddr(), alignment, |
| 2868 | isVolatile: op.getIsVolatile(), |
| 2869 | /*isNonTemporal=*/op.getIsNontemporal(), /*isInvariantGroup=*/false, |
| 2870 | ordering: memorder, syncscope: llvmSyncScope.value_or(u: std::string())); |
| 2871 | if (mlir::Attribute domain = op->getAttr(name: "cir.riscv_nontemporal_domain" )) |
| 2872 | storeOp->setAttr(name: "cir.riscv_nontemporal_domain" , value: domain); |
| 2873 | rewriter.replaceOp(op, newOp: storeOp); |
| 2874 | assert(!cir::MissingFeatures::opLoadStoreTbaa()); |
| 2875 | return mlir::LogicalResult::success(); |
| 2876 | } |
| 2877 | |
| 2878 | static mlir::Type getConstArrayBaseElementType(mlir::Type ty) { |
| 2879 | while (auto arrTy = mlir::dyn_cast<cir::ArrayType>(Val&: ty)) |
| 2880 | ty = arrTy.getElementType(); |
| 2881 | return ty; |
| 2882 | } |
| 2883 | |
| 2884 | static bool isBulkLowerableConstArrayBaseElement(mlir::Type baseElemTy) { |
| 2885 | return mlir::isa<cir::PointerType, cir::IntType, cir::BoolType, |
| 2886 | cir::FPTypeInterface, cir::RecordType>(Val: baseElemTy); |
| 2887 | } |
| 2888 | |
| 2889 | mlir::LogicalResult CIRToLLVMConstantOpLowering::matchAndRewrite( |
| 2890 | cir::ConstantOp op, OpAdaptor adaptor, |
| 2891 | mlir::ConversionPatternRewriter &rewriter) const { |
| 2892 | mlir::Attribute attr = op.getValue(); |
| 2893 | if (!getTypeConverter()->convertType(t: op.getType())) |
| 2894 | return emitTypeConversionError(op, type: op.getType()); |
| 2895 | mlir::ModuleOp moduleOp = op->getParentOfType<mlir::ModuleOp>(); |
| 2896 | |
| 2897 | if (mlir::isa<cir::PoisonAttr>(Val: attr)) { |
| 2898 | rewriter.replaceOpWithNewOp<mlir::LLVM::PoisonOp>( |
| 2899 | op, args: getTypeConverter()->convertType(t: op.getType())); |
| 2900 | return mlir::success(); |
| 2901 | } |
| 2902 | |
| 2903 | if (mlir::isa<cir::UndefAttr>(Val: attr)) { |
| 2904 | rewriter.replaceOpWithNewOp<mlir::LLVM::UndefOp>( |
| 2905 | op, args: getTypeConverter()->convertType(t: op.getType())); |
| 2906 | return mlir::success(); |
| 2907 | } |
| 2908 | |
| 2909 | if (mlir::isa<mlir::IntegerType>(Val: op.getType())) { |
| 2910 | // Verified cir.const operations cannot actually be of these types, but the |
| 2911 | // lowering pass may generate temporary cir.const operations with these |
| 2912 | // types. This is OK since MLIR allows unverified operations to be alive |
| 2913 | // during a pass as long as they don't live past the end of the pass. |
| 2914 | attr = op.getValue(); |
| 2915 | } else if (mlir::isa<cir::BoolType>(Val: op.getType())) { |
| 2916 | int value = mlir::cast<cir::BoolAttr>(Val: op.getValue()).getValue(); |
| 2917 | attr = rewriter.getIntegerAttr(type: typeConverter->convertType(t: op.getType()), |
| 2918 | value); |
| 2919 | } else if (mlir::isa<cir::IntType>(Val: op.getType())) { |
| 2920 | // Lower GlobalViewAttr to llvm.mlir.addressof + llvm.mlir.ptrtoint and |
| 2921 | // BlockAddrDiffAttr to the difference of two llvm.blockaddress. |
| 2922 | if (mlir::isa<cir::GlobalViewAttr, cir::BlockAddrDiffAttr>(Val: op.getValue())) { |
| 2923 | auto newOp = lowerCirAttrAsValue( |
| 2924 | parentLoc: op->getLoc(), parentModule: moduleOp, attr: op.getValue(), rewriter, symbolTables, |
| 2925 | converter: getTypeConverter(), preConversionTypes, blockInfoAddr: &blockInfoAddr); |
| 2926 | rewriter.replaceOp(op, newValues: newOp); |
| 2927 | return mlir::success(); |
| 2928 | } |
| 2929 | |
| 2930 | attr = rewriter.getIntegerAttr( |
| 2931 | type: typeConverter->convertType(t: op.getType()), |
| 2932 | value: mlir::cast<cir::IntAttr>(Val: op.getValue()).getValue()); |
| 2933 | } else if (mlir::isa<cir::FPTypeInterface>(Val: op.getType())) { |
| 2934 | attr = rewriter.getFloatAttr( |
| 2935 | type: typeConverter->convertType(t: op.getType()), |
| 2936 | value: mlir::cast<cir::FPAttr>(Val: op.getValue()).getValue()); |
| 2937 | } else if (mlir::isa<cir::PointerType>(Val: op.getType())) { |
| 2938 | // Optimize with dedicated LLVM op for null pointers. |
| 2939 | if (mlir::isa<cir::ConstPtrAttr>(Val: op.getValue())) { |
| 2940 | if (mlir::cast<cir::ConstPtrAttr>(Val: op.getValue()).isNullValue()) { |
| 2941 | rewriter.replaceOpWithNewOp<mlir::LLVM::ZeroOp>( |
| 2942 | op, args: typeConverter->convertType(t: op.getType())); |
| 2943 | return mlir::success(); |
| 2944 | } |
| 2945 | } |
| 2946 | // Lower GlobalViewAttr and GlobalOffsetAttr to llvm.mlir.addressof, |
| 2947 | // non-null ConstPtrAttr to llvm.inttoptr and BlockAddrInfoAttr to |
| 2948 | // llvm.blockaddress. |
| 2949 | if (mlir::isa<cir::GlobalViewAttr, cir::GlobalOffsetAttr, cir::ConstPtrAttr, |
| 2950 | cir::BlockAddrInfoAttr>(Val: op.getValue())) { |
| 2951 | auto newOp = lowerCirAttrAsValue( |
| 2952 | parentLoc: op->getLoc(), parentModule: moduleOp, attr: op.getValue(), rewriter, symbolTables, |
| 2953 | converter: getTypeConverter(), preConversionTypes, blockInfoAddr: &blockInfoAddr); |
| 2954 | if (!newOp) |
| 2955 | return mlir::failure(); |
| 2956 | rewriter.replaceOp(op, newValues: newOp); |
| 2957 | return mlir::success(); |
| 2958 | } |
| 2959 | attr = op.getValue(); |
| 2960 | } else if (const auto arrTy = mlir::dyn_cast<cir::ArrayType>(Val: op.getType())) { |
| 2961 | // Unlike a global initializer, the constant must keep its type. |
| 2962 | mlir::Attribute value = spellOutSplitTrailingZeros(attr: op.getValue()); |
| 2963 | const auto constArr = mlir::dyn_cast<cir::ConstArrayAttr>(Val&: value); |
| 2964 | if (!constArr && !isa<cir::ZeroAttr, cir::UndefAttr>(Val: value)) |
| 2965 | return op.emitError() << "array does not have a constant initializer" ; |
| 2966 | |
| 2967 | std::optional<mlir::Attribute> denseAttr; |
| 2968 | if (constArr && |
| 2969 | (denseAttr = lowerConstArrayAttr(constArr, symbolTables, converter: typeConverter, |
| 2970 | preConversionTypes, moduleOp))) { |
| 2971 | attr = denseAttr.value(); |
| 2972 | } else { |
| 2973 | const mlir::Value initVal = |
| 2974 | lowerCirAttrAsValue(parentLoc: op->getLoc(), parentModule: moduleOp, attr: value, rewriter, |
| 2975 | symbolTables, converter: typeConverter, preConversionTypes); |
| 2976 | if (!initVal) |
| 2977 | return mlir::failure(); |
| 2978 | rewriter.replaceOp(op, newValues: initVal); |
| 2979 | return mlir::success(); |
| 2980 | } |
| 2981 | } else if (const auto recordAttr = |
| 2982 | mlir::dyn_cast<cir::ConstRecordAttr>(Val: op.getValue())) { |
| 2983 | auto initVal = lowerCirAttrAsValue( |
| 2984 | parentLoc: op->getLoc(), parentModule: moduleOp, attr: spellOutSplitTrailingZeros(attr: recordAttr), |
| 2985 | rewriter, symbolTables, converter: typeConverter, preConversionTypes); |
| 2986 | if (!initVal) |
| 2987 | return mlir::failure(); |
| 2988 | rewriter.replaceOp(op, newValues: initVal); |
| 2989 | return mlir::success(); |
| 2990 | } else if (const auto vecTy = mlir::dyn_cast<cir::VectorType>(Val: op.getType())) { |
| 2991 | mlir::Value initVal = lowerCirAttrAsValue( |
| 2992 | parentLoc: op->getLoc(), parentModule: moduleOp, attr: op.getValue(), rewriter, symbolTables, |
| 2993 | converter: getTypeConverter(), preConversionTypes); |
| 2994 | if (!initVal) |
| 2995 | return mlir::failure(); |
| 2996 | rewriter.replaceOp(op, newValues: initVal); |
| 2997 | return mlir::success(); |
| 2998 | } else if (mlir::isa<cir::RecordType>(Val: op.getType())) { |
| 2999 | if (mlir::isa<cir::ZeroAttr, cir::UndefAttr>(Val: attr)) { |
| 3000 | mlir::Value initVal = |
| 3001 | lowerCirAttrAsValue(parentLoc: op->getLoc(), parentModule: moduleOp, attr, rewriter, |
| 3002 | symbolTables, converter: typeConverter, preConversionTypes); |
| 3003 | if (!initVal) |
| 3004 | return mlir::failure(); |
| 3005 | rewriter.replaceOp(op, newValues: initVal); |
| 3006 | return mlir::success(); |
| 3007 | } |
| 3008 | return op.emitError() << "unsupported lowering for record constant type " |
| 3009 | << op.getType(); |
| 3010 | } else if (auto complexTy = mlir::dyn_cast<cir::ComplexType>(Val: op.getType())) { |
| 3011 | mlir::Type complexElemTy = complexTy.getElementType(); |
| 3012 | mlir::Type complexElemLLVMTy = typeConverter->convertType(t: complexElemTy); |
| 3013 | |
| 3014 | if (auto zeroInitAttr = mlir::dyn_cast<cir::ZeroAttr>(Val: op.getValue())) { |
| 3015 | mlir::TypedAttr zeroAttr = rewriter.getZeroAttr(type: complexElemLLVMTy); |
| 3016 | mlir::ArrayAttr array = rewriter.getArrayAttr(value: {zeroAttr, zeroAttr}); |
| 3017 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( |
| 3018 | op, args: getTypeConverter()->convertType(t: op.getType()), args&: array); |
| 3019 | return mlir::success(); |
| 3020 | } |
| 3021 | |
| 3022 | if (mlir::isa<cir::UndefAttr>(Val: op.getValue())) { |
| 3023 | rewriter.replaceOpWithNewOp<mlir::LLVM::UndefOp>( |
| 3024 | op, args: getTypeConverter()->convertType(t: op.getType())); |
| 3025 | return mlir::success(); |
| 3026 | } |
| 3027 | |
| 3028 | auto complexAttr = mlir::cast<cir::ConstComplexAttr>(Val: op.getValue()); |
| 3029 | |
| 3030 | mlir::Attribute components[2]; |
| 3031 | if (mlir::isa<cir::IntType>(Val: complexElemTy)) { |
| 3032 | components[0] = rewriter.getIntegerAttr( |
| 3033 | type: complexElemLLVMTy, |
| 3034 | value: mlir::cast<cir::IntAttr>(Val: complexAttr.getReal()).getValue()); |
| 3035 | components[1] = rewriter.getIntegerAttr( |
| 3036 | type: complexElemLLVMTy, |
| 3037 | value: mlir::cast<cir::IntAttr>(Val: complexAttr.getImag()).getValue()); |
| 3038 | } else { |
| 3039 | components[0] = rewriter.getFloatAttr( |
| 3040 | type: complexElemLLVMTy, |
| 3041 | value: mlir::cast<cir::FPAttr>(Val: complexAttr.getReal()).getValue()); |
| 3042 | components[1] = rewriter.getFloatAttr( |
| 3043 | type: complexElemLLVMTy, |
| 3044 | value: mlir::cast<cir::FPAttr>(Val: complexAttr.getImag()).getValue()); |
| 3045 | } |
| 3046 | |
| 3047 | attr = rewriter.getArrayAttr(value: components); |
| 3048 | } else { |
| 3049 | return op.emitError() << "unsupported constant type " << op.getType(); |
| 3050 | } |
| 3051 | |
| 3052 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( |
| 3053 | op, args: getTypeConverter()->convertType(t: op.getType()), args&: attr); |
| 3054 | |
| 3055 | return mlir::success(); |
| 3056 | } |
| 3057 | |
| 3058 | static uint64_t getTypeAllocSize(mlir::Type type, mlir::Operation &op) { |
| 3059 | mlir::DataLayout layout(op.getParentOfType<mlir::ModuleOp>()); |
| 3060 | // For LLVM purposes we treat void as u8. |
| 3061 | if (isa<cir::VoidType>(Val: type)) |
| 3062 | type = cir::IntType::get(context: type.getContext(), width: 8, /*isSigned=*/is_signed: false); |
| 3063 | return layout.getTypeAllocSize(t: type); |
| 3064 | } |
| 3065 | |
| 3066 | mlir::LogicalResult CIRToLLVMPrefetchOpLowering::matchAndRewrite( |
| 3067 | cir::PrefetchOp op, OpAdaptor adaptor, |
| 3068 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3069 | rewriter.replaceOpWithNewOp<mlir::LLVM::Prefetch>( |
| 3070 | op, args: adaptor.getAddr(), args: adaptor.getIsWrite(), args: adaptor.getLocality(), |
| 3071 | /*DataCache=*/args: 1); |
| 3072 | return mlir::success(); |
| 3073 | } |
| 3074 | |
| 3075 | mlir::LogicalResult CIRToLLVMPtrDiffOpLowering::matchAndRewrite( |
| 3076 | cir::PtrDiffOp op, OpAdaptor adaptor, |
| 3077 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3078 | mlir::Type pointeeType = op.getLhs().getType().getPointee(); |
| 3079 | if (!convertTypeForMemory(converter: *getTypeConverter(), dataLayout, type: pointeeType)) |
| 3080 | return emitTypeConversionError(op, type: pointeeType); |
| 3081 | |
| 3082 | auto dstTy = mlir::cast<cir::IntType>(Val: op.getType()); |
| 3083 | mlir::Type llvmDstTy = getTypeConverter()->convertType(t: dstTy); |
| 3084 | |
| 3085 | auto lhs = mlir::LLVM::PtrToIntOp::create(builder&: rewriter, location: op.getLoc(), res: llvmDstTy, |
| 3086 | arg: adaptor.getLhs()); |
| 3087 | auto rhs = mlir::LLVM::PtrToIntOp::create(builder&: rewriter, location: op.getLoc(), res: llvmDstTy, |
| 3088 | arg: adaptor.getRhs()); |
| 3089 | |
| 3090 | auto diff = |
| 3091 | mlir::LLVM::SubOp::create(builder&: rewriter, location: op.getLoc(), res: llvmDstTy, lhs, rhs); |
| 3092 | |
| 3093 | assert(!cir::MissingFeatures::llvmLoweringPtrDiffConsidersPointee()); |
| 3094 | uint64_t typeSize = getTypeAllocSize(type: pointeeType, op&: *op); |
| 3095 | |
| 3096 | // Avoid silly division by 1. |
| 3097 | mlir::Value resultVal = diff.getResult(); |
| 3098 | if (typeSize != 1) { |
| 3099 | auto typeSizeVal = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 3100 | type: llvmDstTy, value: typeSize); |
| 3101 | |
| 3102 | if (dstTy.isUnsigned()) { |
| 3103 | auto uDiv = |
| 3104 | mlir::LLVM::UDivOp::create(builder&: rewriter, location: op.getLoc(), lhs: diff, rhs: typeSizeVal); |
| 3105 | uDiv.setIsExact(true); |
| 3106 | resultVal = uDiv.getResult(); |
| 3107 | } else { |
| 3108 | auto sDiv = |
| 3109 | mlir::LLVM::SDivOp::create(builder&: rewriter, location: op.getLoc(), lhs: diff, rhs: typeSizeVal); |
| 3110 | sDiv.setIsExact(true); |
| 3111 | resultVal = sDiv.getResult(); |
| 3112 | } |
| 3113 | } |
| 3114 | rewriter.replaceOp(op, newValues: resultVal); |
| 3115 | return mlir::success(); |
| 3116 | } |
| 3117 | |
| 3118 | mlir::LogicalResult CIRToLLVMExpectOpLowering::matchAndRewrite( |
| 3119 | cir::ExpectOp op, OpAdaptor adaptor, |
| 3120 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3121 | // TODO(cir): do not generate LLVM intrinsics under -O0 |
| 3122 | assert(!cir::MissingFeatures::optInfoAttr()); |
| 3123 | |
| 3124 | std::optional<llvm::APFloat> prob = op.getProb(); |
| 3125 | if (prob) |
| 3126 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExpectWithProbabilityOp>( |
| 3127 | op, args: adaptor.getVal(), args: adaptor.getExpected(), args&: prob.value()); |
| 3128 | else |
| 3129 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExpectOp>(op, args: adaptor.getVal(), |
| 3130 | args: adaptor.getExpected()); |
| 3131 | return mlir::success(); |
| 3132 | } |
| 3133 | |
| 3134 | /// Return true for discardable CIR markers handled explicitly below. |
| 3135 | static bool isHandledDiscardableFuncAttr(mlir::NamedAttribute attr) { |
| 3136 | return attr.getName() == CIRDialect::getNoReturnAttrName() || |
| 3137 | attr.getName() == CIRDialect::getNoUnwindAttrName() || |
| 3138 | attr.getName() == CIRDialect::getWillReturnAttrName() || |
| 3139 | attr.getName() == CIRDialect::getStrictFPAttrName() || |
| 3140 | attr.getName() == CIRDialect::getNoRecurseAttrName() || |
| 3141 | attr.getName() == CIRDialect::getMustProgressAttrName() || |
| 3142 | attr.getName() == CIRDialect::getNoBuiltinAttrName() || |
| 3143 | attr.getName() == CIRDialect::getOptNoneAttrName() || |
| 3144 | attr.getName() == CIRDialect::getNakedAttrName() || |
| 3145 | attr.getName() == CIRDialect::getSYCLModuleIdAttrName() || |
| 3146 | attr.getName() == CIRDialect::getDontCallErrorAttrName() || |
| 3147 | attr.getName() == CIRDialect::getDontCallWarnAttrName() || |
| 3148 | attr.getName() == CIRDialect::getRuntimeFuncAttrName(); |
| 3149 | } |
| 3150 | |
| 3151 | /// Lower `cir.func` attributes for an `LLVMFuncOp` or `LLVM::AliasOp`. |
| 3152 | /// Drop attributes populated by the destination op builder. If |
| 3153 | /// `includeFunctionOnlyAttrs` is false, also omit attributes that are only |
| 3154 | /// valid on functions. Fails when an argument attribute's pointee type has |
| 3155 | /// no memory representation. |
| 3156 | mlir::LogicalResult CIRToLLVMFuncOpLowering::lowerFuncAttributes( |
| 3157 | cir::FuncOp func, bool includeFunctionOnlyAttrs, |
| 3158 | SmallVectorImpl<mlir::NamedAttribute> &result) const { |
| 3159 | FunctionMetadataLowering metadataLowering(func.getContext()); |
| 3160 | for (mlir::NamedAttribute attr : func->getDiscardableAttrs()) { |
| 3161 | if (isHandledDiscardableFuncAttr(attr)) |
| 3162 | continue; |
| 3163 | if (metadataLowering.lower(attr, includeFunctionOnlyAttrs)) |
| 3164 | continue; |
| 3165 | assert(!cir::MissingFeatures::opFuncExtraAttrs()); |
| 3166 | result.push_back(Elt: attr); |
| 3167 | } |
| 3168 | |
| 3169 | if (mlir::StringAttr visibility = func.getSymVisibilityAttr()) |
| 3170 | result.emplace_back(Args: func.getSymVisibilityAttrName(), Args&: visibility); |
| 3171 | |
| 3172 | if (includeFunctionOnlyAttrs) { |
| 3173 | if (mlir::ArrayAttr argAttrs = func.getArgAttrsAttr()) { |
| 3174 | std::optional<mlir::ArrayAttr> lowered = convertTypedArgAttrs( |
| 3175 | argAttrs, converter: *getTypeConverter(), dataLayout, ctx: getContext()); |
| 3176 | if (!lowered) |
| 3177 | return func.emitError() << "NYI: lowering a byval/sret/byref " |
| 3178 | "argument whose pointee type has no " |
| 3179 | "memory representation" ; |
| 3180 | result.emplace_back(Args: func.getArgAttrsAttrName(), Args&: *lowered); |
| 3181 | } |
| 3182 | if (mlir::ArrayAttr resAttrs = func.getResAttrsAttr()) |
| 3183 | result.emplace_back(Args: func.getResAttrsAttrName(), Args&: resAttrs); |
| 3184 | } |
| 3185 | |
| 3186 | if (includeFunctionOnlyAttrs) |
| 3187 | metadataLowering.appendAttrs(result); |
| 3188 | |
| 3189 | return mlir::success(); |
| 3190 | } |
| 3191 | |
| 3192 | mlir::LogicalResult CIRToLLVMFuncOpLowering::matchAndRewriteAlias( |
| 3193 | cir::FuncOp op, llvm::StringRef aliasee, mlir::Type ty, OpAdaptor adaptor, |
| 3194 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3195 | SmallVector<mlir::NamedAttribute, 4> attributes; |
| 3196 | if (mlir::failed(Result: lowerFuncAttributes(func: op, /*includeFunctionOnlyAttrs=*/false, |
| 3197 | result&: attributes))) |
| 3198 | return mlir::failure(); |
| 3199 | |
| 3200 | mlir::Location loc = op.getLoc(); |
| 3201 | // Create the alias before replacing: without pattern rollback the |
| 3202 | // replace erases the root op immediately. The driver does not |
| 3203 | // invalidate the cached symbol table either, so build it before the |
| 3204 | // replacement exists and maintain it explicitly below. |
| 3205 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3206 | symbolTables.getSymbolTable(op: modOp); |
| 3207 | auto aliasOp = mlir::LLVM::AliasOp::create( |
| 3208 | builder&: rewriter, location: loc, type: ty, linkage: convertLinkage(linkage: op.getLinkage()), name: op.getName(), |
| 3209 | dsoLocal: op.getDsoLocal(), tls_mode: mlir::LLVM::ThreadLocalMode::NotThreadLocal, |
| 3210 | attrs: attributes); |
| 3211 | aliasOp.setVisibility_( |
| 3212 | lowerCIRVisibilityToLLVMVisibility(visibilityKind: op.getGlobalVisibility())); |
| 3213 | if (op.getUnnamedAddr()) |
| 3214 | aliasOp.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Global); |
| 3215 | { |
| 3216 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 3217 | rewriter.createBlock(parent: &aliasOp.getInitializerRegion()); |
| 3218 | cir::CIRDataLayout cirDataLayout( |
| 3219 | aliasOp->getParentOfType<mlir::ModuleOp>()); |
| 3220 | mlir::Type ptrTy = mlir::LLVM::LLVMPointerType::get( |
| 3221 | context: ty.getContext(), addressSpace: cirDataLayout.getProgramAddressSpace()); |
| 3222 | auto addrOp = |
| 3223 | mlir::LLVM::AddressOfOp::create(builder&: rewriter, location: loc, res: ptrTy, global_name: aliasee); |
| 3224 | mlir::LLVM::ReturnOp::create(builder&: rewriter, location: loc, arg: addrOp); |
| 3225 | } |
| 3226 | symbolTables.getSymbolTable(op: modOp).remove(op); |
| 3227 | rewriter.replaceOp(op, newOp: aliasOp); |
| 3228 | symbolTables.getSymbolTable(op: modOp).insert(symbol: aliasOp); |
| 3229 | |
| 3230 | return mlir::success(); |
| 3231 | } |
| 3232 | |
| 3233 | mlir::LogicalResult CIRToLLVMFuncOpLowering::matchAndRewrite( |
| 3234 | cir::FuncOp op, OpAdaptor adaptor, |
| 3235 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3236 | |
| 3237 | cir::FuncType fnType = op.getFunctionType(); |
| 3238 | bool isDsoLocal = op.getDsoLocal(); |
| 3239 | mlir::TypeConverter::SignatureConversion signatureConversion( |
| 3240 | fnType.getNumInputs()); |
| 3241 | |
| 3242 | for (const auto &argType : llvm::enumerate(First: fnType.getInputs())) { |
| 3243 | mlir::Type convertedType = typeConverter->convertType(t: argType.value()); |
| 3244 | if (!convertedType) |
| 3245 | return emitTypeConversionError(op, type: argType.value()); |
| 3246 | signatureConversion.addInputs(origInputNo: argType.index(), types: convertedType); |
| 3247 | } |
| 3248 | |
| 3249 | mlir::Type resultType = |
| 3250 | getTypeConverter()->convertType(t: fnType.getReturnType()); |
| 3251 | if (!resultType) |
| 3252 | return emitTypeConversionError(op, type: fnType.getReturnType()); |
| 3253 | |
| 3254 | // Create the LLVM function operation. |
| 3255 | mlir::Type llvmFnTy = mlir::LLVM::LLVMFunctionType::get( |
| 3256 | result: resultType, arguments: signatureConversion.getConvertedTypes(), |
| 3257 | /*isVarArg=*/fnType.isVarArg()); |
| 3258 | |
| 3259 | // If this is an alias, it needs to be lowered to llvm::AliasOp. |
| 3260 | if (std::optional<llvm::StringRef> aliasee = op.getAliasee()) |
| 3261 | return matchAndRewriteAlias(op, aliasee: *aliasee, ty: llvmFnTy, adaptor, rewriter); |
| 3262 | |
| 3263 | // LLVMFuncOp expects a single FileLine Location instead of a fused |
| 3264 | // location. |
| 3265 | mlir::Location loc = op.getLoc(); |
| 3266 | if (mlir::FusedLoc fusedLoc = mlir::dyn_cast<mlir::FusedLoc>(Val&: loc)) |
| 3267 | loc = fusedLoc.getLocations()[0]; |
| 3268 | assert((mlir::isa<mlir::FileLineColLoc>(loc) || |
| 3269 | mlir::isa<mlir::UnknownLoc>(loc)) && |
| 3270 | "expected single location or unknown location here" ); |
| 3271 | |
| 3272 | mlir::LLVM::Linkage linkage = convertLinkage(linkage: op.getLinkage()); |
| 3273 | mlir::LLVM::CConv cconv = convertCallingConv(callingConv: op.getCallingConv()); |
| 3274 | SmallVector<mlir::NamedAttribute, 4> attributes; |
| 3275 | if (mlir::failed(Result: lowerFuncAttributes(func: op, /*includeFunctionOnlyAttrs=*/true, |
| 3276 | result&: attributes))) |
| 3277 | return mlir::failure(); |
| 3278 | |
| 3279 | // Convert the body types here: this mutates the op, so it runs after the |
| 3280 | // last fallible read above and before any op creation below. A failure |
| 3281 | // leaves the body unmoved for the driver to report. |
| 3282 | if (failed(Result: rewriter.convertRegionTypes(region: &op.getBody(), converter: *typeConverter, |
| 3283 | entryConversion: &signatureConversion))) |
| 3284 | return mlir::failure(); |
| 3285 | |
| 3286 | mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, builder&: rewriter); |
| 3287 | |
| 3288 | // The driver does not invalidate the cached symbol table, so build it |
| 3289 | // before the replacement exists and maintain it explicitly below. |
| 3290 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3291 | symbolTables.getSymbolTable(op: modOp); |
| 3292 | mlir::LLVM::LLVMFuncOp fn = mlir::LLVM::LLVMFuncOp::create( |
| 3293 | builder&: rewriter, location: loc, name: op.getName(), type: llvmFnTy, linkage, dsoLocal: isDsoLocal, cconv, |
| 3294 | comdat: comdatAttr, attrs: attributes); |
| 3295 | |
| 3296 | assert(!cir::MissingFeatures::opFuncMultipleReturnVals()); |
| 3297 | |
| 3298 | if (op.getUnnamedAddr()) |
| 3299 | fn.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Global); |
| 3300 | |
| 3301 | if (cir::MemoryEffectsAttr effects = op.getMemoryEffectsAttr()) |
| 3302 | fn.setMemoryEffectsAttr(convertMemoryEffects(ctx: fn.getContext(), effects)); |
| 3303 | |
| 3304 | // As for calls, either cannot-unwind fact maps to LLVM's nounwind. |
| 3305 | if (op->hasAttr(name: CIRDialect::getNoUnwindAttrName()) || |
| 3306 | op->hasAttr(name: CIRDialect::getNoThrowAttrName())) |
| 3307 | fn.setNoUnwind(true); |
| 3308 | if (op->hasAttr(name: CIRDialect::getWillReturnAttrName())) |
| 3309 | fn.setWillReturn(true); |
| 3310 | |
| 3311 | if (op->hasAttr(name: CIRDialect::getNoReturnAttrName())) |
| 3312 | fn.setNoreturn(true); |
| 3313 | |
| 3314 | if (op->hasAttr(name: CIRDialect::getOptNoneAttrName())) |
| 3315 | fn.setOptimizeNone(true); |
| 3316 | |
| 3317 | if (std::optional<cir::UnwindTableKind> uwtableKind = op.getUwtable()) |
| 3318 | fn.setUwtableKindAttr(mlir::LLVM::UWTableKindAttr::get( |
| 3319 | context: fn.getContext(), uwtableKind: convertUWTableKind(kind: *uwtableKind))); |
| 3320 | |
| 3321 | if (std::optional<cir::FramePointerKind> framePointerKind = |
| 3322 | op.getFramePointer()) |
| 3323 | fn.setFramePointerAttr(mlir::LLVM::FramePointerKindAttr::get( |
| 3324 | context: fn.getContext(), framePointerKind: convertFramePointerKind(kind: *framePointerKind))); |
| 3325 | |
| 3326 | // Function attributes with no dedicated field on the LLVM dialect's |
| 3327 | // LLVMFuncOp are routed through the `passthrough` array. The MLIR LLVM IR |
| 3328 | // translator forwards `passthrough` entries to LLVM IR as function |
| 3329 | // attributes. |
| 3330 | SmallVector<mlir::Attribute> passthrough; |
| 3331 | for (llvm::StringRef flagAttr : |
| 3332 | {CIRDialect::getStrictFPAttrName(), CIRDialect::getNoRecurseAttrName(), |
| 3333 | CIRDialect::getMustProgressAttrName(), |
| 3334 | CIRDialect::getNoBuiltinAttrName(), CIRDialect::getNakedAttrName()}) |
| 3335 | if (op->hasAttr(name: flagAttr)) |
| 3336 | passthrough.push_back(Elt: rewriter.getStringAttr(bytes: flagAttr)); |
| 3337 | |
| 3338 | if (auto moduleId = op->getAttrOfType<mlir::StringAttr>( |
| 3339 | name: CIRDialect::getSYCLModuleIdAttrName())) |
| 3340 | passthrough.push_back(Elt: rewriter.getArrayAttr( |
| 3341 | value: {rewriter.getStringAttr(bytes: CIRDialect::getSYCLModuleIdAttrName()), |
| 3342 | moduleId})); |
| 3343 | |
| 3344 | for (llvm::StringRef dontCallAttr : {CIRDialect::getDontCallErrorAttrName(), |
| 3345 | CIRDialect::getDontCallWarnAttrName()}) |
| 3346 | if (auto diagnostic = op->getAttrOfType<mlir::StringAttr>(name: dontCallAttr)) |
| 3347 | passthrough.push_back(Elt: rewriter.getArrayAttr( |
| 3348 | value: {rewriter.getStringAttr(bytes: dontCallAttr), diagnostic})); |
| 3349 | |
| 3350 | // Every function but a runtime function gets the TU-wide default function |
| 3351 | // attributes. |
| 3352 | if (!op->hasAttr(name: CIRDialect::getRuntimeFuncAttrName())) { |
| 3353 | if (auto langOpts = op->getParentOfType<mlir::ModuleOp>() |
| 3354 | ->getAttrOfType<cir::LoweringOptionsAttr>( |
| 3355 | name: CIRDialect::getLoweringOptionsAttrName())) { |
| 3356 | // A strictfp function may use a pragma to observe FP exceptions locally, |
| 3357 | // so the TU-wide default does not apply to it. |
| 3358 | if (langOpts.getNoTrappingMath() && |
| 3359 | !op->hasAttr(name: CIRDialect::getStrictFPAttrName())) |
| 3360 | passthrough.push_back( |
| 3361 | Elt: rewriter.getStrArrayAttr(values: {"no-trapping-math" , "true" })); |
| 3362 | passthrough.push_back(Elt: rewriter.getStrArrayAttr( |
| 3363 | values: {"stack-protector-buffer-size" , |
| 3364 | llvm::utostr(X: langOpts.getSspBufferSize())})); |
| 3365 | } |
| 3366 | } |
| 3367 | |
| 3368 | if (op.getCoroutineKind()) |
| 3369 | passthrough.push_back(Elt: rewriter.getStringAttr(bytes: "presplitcoroutine" )); |
| 3370 | if (!passthrough.empty()) |
| 3371 | fn.setPassthroughAttr(rewriter.getArrayAttr(value: passthrough)); |
| 3372 | |
| 3373 | if (std::optional<cir::InlineKind> inlineKind = op.getInlineKind()) { |
| 3374 | fn.setNoInline(*inlineKind == cir::InlineKind::NoInline); |
| 3375 | fn.setInlineHint(*inlineKind == cir::InlineKind::InlineHint); |
| 3376 | fn.setAlwaysInline(*inlineKind == cir::InlineKind::AlwaysInline); |
| 3377 | } |
| 3378 | |
| 3379 | if (std::optional<llvm::StringRef> personality = op.getPersonality()) |
| 3380 | fn.setPersonality(*personality); |
| 3381 | |
| 3382 | fn.setVisibility_( |
| 3383 | lowerCIRVisibilityToLLVMVisibility(visibilityKind: op.getGlobalVisibility())); |
| 3384 | |
| 3385 | fn.setAlignment(op.getAlignment()); |
| 3386 | assert(!MissingFeatures::opFuncPreferredAlignment()); |
| 3387 | |
| 3388 | symbolTables.getSymbolTable(op: modOp).remove(op); |
| 3389 | rewriter.inlineRegionBefore(region&: op.getBody(), parent&: fn.getBody(), before: fn.end()); |
| 3390 | |
| 3391 | rewriter.eraseOp(op); |
| 3392 | symbolTables.getSymbolTable(op: modOp).insert(symbol: fn); |
| 3393 | |
| 3394 | return mlir::LogicalResult::success(); |
| 3395 | } |
| 3396 | |
| 3397 | mlir::LogicalResult CIRToLLVMGetGlobalOpLowering::matchAndRewrite( |
| 3398 | cir::GetGlobalOp op, OpAdaptor adaptor, |
| 3399 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3400 | // FIXME(cir): Premature DCE to avoid lowering stuff we're not using. |
| 3401 | // CIRGen should mitigate this and not emit the get_global. |
| 3402 | if (op->getUses().empty()) { |
| 3403 | rewriter.eraseOp(op); |
| 3404 | return mlir::success(); |
| 3405 | } |
| 3406 | |
| 3407 | mlir::Type type = getTypeConverter()->convertType(t: op.getType()); |
| 3408 | if (!type) |
| 3409 | return emitTypeConversionError(op, type: op.getType()); |
| 3410 | mlir::Operation *newop = mlir::LLVM::AddressOfOp::create( |
| 3411 | builder&: rewriter, location: op.getLoc(), res: type, global_name: op.getName()); |
| 3412 | |
| 3413 | if (op.getTls()) { |
| 3414 | // Handle access to TLS via intrinsic. |
| 3415 | newop = mlir::LLVM::ThreadlocalAddressOp::create(builder&: rewriter, location: op.getLoc(), |
| 3416 | res: type, global: newop->getResult(idx: 0)); |
| 3417 | } |
| 3418 | |
| 3419 | rewriter.replaceOp(op, newOp: newop); |
| 3420 | return mlir::success(); |
| 3421 | } |
| 3422 | |
| 3423 | llvm::SmallVector<mlir::NamedAttribute> |
| 3424 | CIRToLLVMGlobalOpLowering::lowerGlobalAttributes( |
| 3425 | cir::GlobalOp op, mlir::ConversionPatternRewriter &rewriter) const { |
| 3426 | SmallVector<mlir::NamedAttribute> attributes; |
| 3427 | |
| 3428 | if (mlir::StringAttr sectionAttr = op.getSectionAttr()) |
| 3429 | attributes.push_back(Elt: rewriter.getNamedAttr(name: "section" , val: sectionAttr)); |
| 3430 | |
| 3431 | mlir::LLVM::VisibilityAttr visibility = mlir::LLVM::VisibilityAttr::get( |
| 3432 | context: getContext(), |
| 3433 | val: lowerCIRVisibilityToLLVMVisibility(visibilityKind: op.getGlobalVisibility())); |
| 3434 | attributes.push_back(Elt: rewriter.getNamedAttr(name: "visibility_" , val: visibility)); |
| 3435 | |
| 3436 | if (op.getUnnamedAddr()) |
| 3437 | attributes.push_back(Elt: rewriter.getNamedAttr( |
| 3438 | name: "unnamed_addr" , val: mlir::LLVM::UnnamedAddrAttr::get( |
| 3439 | context: getContext(), val: mlir::LLVM::UnnamedAddr::Global))); |
| 3440 | |
| 3441 | if (op->getAttr(name: CUDAExternallyInitializedAttr::getMnemonic())) |
| 3442 | attributes.push_back(Elt: rewriter.getNamedAttr(name: "externally_initialized" , |
| 3443 | val: rewriter.getUnitAttr())); |
| 3444 | |
| 3445 | return attributes; |
| 3446 | } |
| 3447 | |
| 3448 | static mlir::LLVM::ThreadLocalMode |
| 3449 | convertTlsModelAttrToLLVM(TLSModelAttr attr) { |
| 3450 | // assert that we can just static-cast these. |
| 3451 | #define CHECK_ENUM(CIR, LLVM_VAL) \ |
| 3452 | static_assert(static_cast<unsigned>(TLSModel::CIR) == \ |
| 3453 | static_cast<unsigned>(mlir::LLVM::ThreadLocalMode::LLVM_VAL)) |
| 3454 | CHECK_ENUM(GeneralDynamic, GeneralDynamic); |
| 3455 | CHECK_ENUM(LocalDynamic, LocalDynamic); |
| 3456 | CHECK_ENUM(InitialExec, InitialExec); |
| 3457 | CHECK_ENUM(LocalExec, LocalExec); |
| 3458 | #undef CHECK_ENUM |
| 3459 | |
| 3460 | if (!attr) |
| 3461 | return mlir::LLVM::ThreadLocalMode::NotThreadLocal; |
| 3462 | |
| 3463 | return static_cast<mlir::LLVM::ThreadLocalMode>(attr.getValue()); |
| 3464 | } |
| 3465 | |
| 3466 | /// Create the region-initialized LLVM global beside the CIR op and update |
| 3467 | /// insertion point to the end of the initializer block. Returns the new op, |
| 3468 | /// or null on failure. The caller owns the replacement: it runs the |
| 3469 | /// initializer visit with the old op still in the table, then swaps the |
| 3470 | /// table and erases the old op on success. |
| 3471 | mlir::LLVM::GlobalOp |
| 3472 | CIRToLLVMGlobalOpLowering::setupRegionInitializedLLVMGlobalOp( |
| 3473 | cir::GlobalOp op, mlir::ConversionPatternRewriter &rewriter) const { |
| 3474 | mlir::Type llvmType = |
| 3475 | convertTypeForMemory(converter: *getTypeConverter(), dataLayout, type: op.getSymType()); |
| 3476 | |
| 3477 | // Keep the global's type in sync with the value built by CIRAttrToValue: a |
| 3478 | // flexible array member initializer requires an oversized anonymous struct. |
| 3479 | mlir::Type failedType = op.getSymType(); |
| 3480 | if (std::optional<mlir::Attribute> init = op.getInitialValue()) |
| 3481 | llvmType = adjustGlobalTypeForInit(llvmType, init: *init, converter: *getTypeConverter(), |
| 3482 | dataLayout, failedType: &failedType); |
| 3483 | if (!llvmType) { |
| 3484 | emitTypeConversionError(op, type: failedType); |
| 3485 | return {}; |
| 3486 | } |
| 3487 | |
| 3488 | // FIXME: These default values are placeholders until the the equivalent |
| 3489 | // attributes are available on cir.global ops. This duplicates code |
| 3490 | // in CIRToLLVMGlobalOpLowering::matchAndRewrite() but that will go |
| 3491 | // away when the placeholders are no longer needed. |
| 3492 | const bool isConst = op.getConstant(); |
| 3493 | unsigned addrSpace = 0; |
| 3494 | if (auto targetAS = mlir::dyn_cast_if_present<cir::TargetAddressSpaceAttr>( |
| 3495 | Val: op.getAddrSpaceAttr())) |
| 3496 | addrSpace = targetAS.getValue(); |
| 3497 | const bool isDsoLocal = op.getDsoLocal(); |
| 3498 | mlir::LLVM::ThreadLocalMode threadLocalMode = |
| 3499 | convertTlsModelAttrToLLVM(attr: op.getTlsModelAttr()); |
| 3500 | const uint64_t alignment = op.getAlignment().value_or(u: 0); |
| 3501 | const mlir::LLVM::Linkage linkage = convertLinkage(linkage: op.getLinkage()); |
| 3502 | const StringRef symbol = op.getSymName(); |
| 3503 | |
| 3504 | SmallVector<mlir::NamedAttribute> attributes = |
| 3505 | lowerGlobalAttributes(op, rewriter); |
| 3506 | |
| 3507 | mlir::LLVM::GlobalOp newGlobalOp = mlir::LLVM::GlobalOp::create( |
| 3508 | builder&: rewriter, location: op.getLoc(), type: llvmType, isConstant: isConst, linkage, name: symbol, value: nullptr, |
| 3509 | alignment, addrSpace, dsoLocal: isDsoLocal, tls_mode: threadLocalMode, comdat: mlir::SymbolRefAttr{}, |
| 3510 | attrs: attributes); |
| 3511 | rewriter.createBlock(parent: &newGlobalOp.getRegion()); |
| 3512 | rewriter.setInsertionPointToEnd(newGlobalOp.getInitializerBlock()); |
| 3513 | return newGlobalOp; |
| 3514 | } |
| 3515 | |
| 3516 | mlir::LogicalResult |
| 3517 | CIRToLLVMGlobalOpLowering::matchAndRewriteRegionInitializedGlobal( |
| 3518 | cir::GlobalOp op, mlir::Attribute init, |
| 3519 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3520 | // TODO: Generalize this handling when more types are needed here. |
| 3521 | assert((isa<cir::BlockAddrDiffAttr, cir::BlockAddrInfoAttr, |
| 3522 | cir::ConstArrayAttr, cir::ConstRecordAttr, cir::ConstVectorAttr, |
| 3523 | cir::ConstPtrAttr, cir::ConstComplexAttr, cir::GlobalOffsetAttr, |
| 3524 | cir::GlobalViewAttr, cir::TypeInfoAttr, cir::UndefAttr, |
| 3525 | cir::PoisonAttr, cir::VTableAttr, cir::ZeroAttr>(init))); |
| 3526 | |
| 3527 | // TODO(cir): once LLVM's dialect has proper equivalent attributes this |
| 3528 | // should be updated. For now, we use a custom op to initialize globals |
| 3529 | // to the appropriate value. |
| 3530 | const mlir::Location loc = op.getLoc(); |
| 3531 | mlir::ModuleOp moduleOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3532 | // Build the cached table before the replacement exists, so the swap below |
| 3533 | // drops a real entry instead of depending on duplicate tolerance. |
| 3534 | auto &symbolTable = symbolTables.getSymbolTable(op: moduleOp); |
| 3535 | // Create the replacement beside the root op, not via replaceOp: the table |
| 3536 | // swap below and the visit both need the old op alive. |
| 3537 | rewriter.setInsertionPoint(op); |
| 3538 | mlir::LLVM::GlobalOp newGlobalOp = |
| 3539 | setupRegionInitializedLLVMGlobalOp(op, rewriter); |
| 3540 | if (!newGlobalOp) |
| 3541 | return mlir::failure(); |
| 3542 | // Keep the old op in the table during the visit so self-references |
| 3543 | // resolve to the CIR type, not the possibly adjusted LLVM type. |
| 3544 | // Pass blockInfoAddr so that block address initializers (either as the whole |
| 3545 | // initializer or nested inside an aggregate) can be resolved by the |
| 3546 | // BlockAddrInfoAttr visitor. |
| 3547 | CIRAttrToValue valueConverter(loc, moduleOp, rewriter, symbolTables, |
| 3548 | typeConverter, preConversionTypes, |
| 3549 | &blockInfoAddr); |
| 3550 | mlir::Value value = valueConverter.visit(attr: init); |
| 3551 | if (!value) { |
| 3552 | // Drop the speculative replacement; the table still holds the live |
| 3553 | // root op for the driver to report on. Everything created so far |
| 3554 | // lives inside the new global's region, so erasing it restores the |
| 3555 | // module exactly: module-level ops must wait for a successful visit |
| 3556 | // (see the comdat below). |
| 3557 | rewriter.eraseOp(op: newGlobalOp); |
| 3558 | return mlir::failure(); |
| 3559 | } |
| 3560 | mlir::LLVM::ReturnOp::create(builder&: rewriter, location: loc, arg: value); |
| 3561 | // Attach comdat only after the visit succeeds, so the failure path |
| 3562 | // above only has the new global itself to drop. |
| 3563 | if (mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, builder&: rewriter)) |
| 3564 | newGlobalOp.setComdatAttr(comdatAttr); |
| 3565 | symbolTable.remove(op); |
| 3566 | symbolTable.insert(symbol: newGlobalOp); |
| 3567 | rewriter.eraseOp(op); |
| 3568 | return mlir::success(); |
| 3569 | } |
| 3570 | |
| 3571 | // Replace a CIR global with an LLVM global, maintaining the cached symbol |
| 3572 | // table. The table entry is dropped before creation so the replacement |
| 3573 | // reuses the name. |
| 3574 | static mlir::LLVM::GlobalOp |
| 3575 | replaceCirGlobal(mlir::SymbolTableCollection &symbolTables, cir::GlobalOp op, |
| 3576 | mlir::Type llvmType, bool isConst, mlir::LLVM::Linkage linkage, |
| 3577 | llvm::StringRef symbol, mlir::Attribute init, |
| 3578 | uint64_t alignment, unsigned addrSpace, bool isDsoLocal, |
| 3579 | mlir::LLVM::ThreadLocalMode threadLocalMode, |
| 3580 | mlir::SymbolRefAttr comdatAttr, |
| 3581 | llvm::ArrayRef<mlir::NamedAttribute> attributes, |
| 3582 | mlir::ConversionPatternRewriter &rewriter) { |
| 3583 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3584 | symbolTables.getSymbolTable(op: modOp).remove(op); |
| 3585 | auto newGlobalOp = rewriter.replaceOpWithNewOp<mlir::LLVM::GlobalOp>( |
| 3586 | op, args&: llvmType, args&: isConst, args&: linkage, args&: symbol, args&: init, args&: alignment, args&: addrSpace, |
| 3587 | args&: isDsoLocal, args&: threadLocalMode, args&: comdatAttr, args&: attributes); |
| 3588 | symbolTables.getSymbolTable(op: modOp).insert(symbol: newGlobalOp); |
| 3589 | return newGlobalOp; |
| 3590 | } |
| 3591 | |
| 3592 | mlir::LogicalResult CIRToLLVMGlobalOpLowering::matchAndRewrite( |
| 3593 | cir::GlobalOp op, OpAdaptor adaptor, |
| 3594 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3595 | if (mlir::isa_and_present<cir::LangAddressSpaceAttr>(Val: op.getAddrSpaceAttr())) |
| 3596 | return op.emitError() |
| 3597 | << "cannot lower a global with a language address space" ; |
| 3598 | |
| 3599 | // If this global requires non-trivial initialization or destruction, |
| 3600 | // that needs to be moved to runtime handlers during LoweringPrepare. |
| 3601 | if (!op.getCtorRegion().empty() || !op.getDtorRegion().empty()) |
| 3602 | return op.emitError() << "GlobalOp ctor and dtor regions should be removed " |
| 3603 | "in LoweringPrepare" ; |
| 3604 | |
| 3605 | std::optional<mlir::Attribute> init = op.getInitialValue(); |
| 3606 | |
| 3607 | // Fetch required values to create LLVM op. |
| 3608 | const mlir::Type cirSymType = op.getSymType(); |
| 3609 | |
| 3610 | // This is the LLVM dialect type. |
| 3611 | mlir::Type llvmType = |
| 3612 | convertTypeForMemory(converter: *getTypeConverter(), dataLayout, type: cirSymType); |
| 3613 | if (!llvmType) |
| 3614 | return op.emitError() |
| 3615 | << "NYI: lowering global of a type with no memory representation" ; |
| 3616 | |
| 3617 | // A flexible array member initializer makes the constant larger than the |
| 3618 | // record's declared type, so the global must use an oversized anonymous |
| 3619 | // struct instead. |
| 3620 | mlir::Type failedType = cirSymType; |
| 3621 | if (init.has_value()) |
| 3622 | llvmType = adjustGlobalTypeForInit(llvmType, init: *init, converter: *getTypeConverter(), |
| 3623 | dataLayout, failedType: &failedType); |
| 3624 | if (!llvmType) |
| 3625 | return emitTypeConversionError(op, type: failedType); |
| 3626 | |
| 3627 | // FIXME: These default values are placeholders until the the equivalent |
| 3628 | // attributes are available on cir.global ops. |
| 3629 | const bool isConst = op.getConstant(); |
| 3630 | unsigned addrSpace = 0; |
| 3631 | if (auto targetAS = mlir::dyn_cast_if_present<cir::TargetAddressSpaceAttr>( |
| 3632 | Val: op.getAddrSpaceAttr())) |
| 3633 | addrSpace = targetAS.getValue(); |
| 3634 | const bool isDsoLocal = op.getDsoLocal(); |
| 3635 | mlir::LLVM::ThreadLocalMode threadLocalMode = |
| 3636 | convertTlsModelAttrToLLVM(attr: op.getTlsModelAttr()); |
| 3637 | const uint64_t alignment = op.getAlignment().value_or(u: 0); |
| 3638 | const mlir::LLVM::Linkage linkage = convertLinkage(linkage: op.getLinkage()); |
| 3639 | const StringRef symbol = op.getSymName(); |
| 3640 | SmallVector<mlir::NamedAttribute> attributes = |
| 3641 | lowerGlobalAttributes(op, rewriter); |
| 3642 | |
| 3643 | // If this is a variable alias, lower it to llvm.mlir.alias. |
| 3644 | if (std::optional<llvm::StringRef> aliasee = op.getAliasee()) { |
| 3645 | mlir::Location loc = op.getLoc(); |
| 3646 | // Create the alias before replacing: without pattern rollback the |
| 3647 | // replace erases the root op immediately. The driver does not |
| 3648 | // invalidate the cached symbol table either, so build it before the |
| 3649 | // replacement exists and maintain it explicitly below. |
| 3650 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3651 | symbolTables.getSymbolTable(op: modOp); |
| 3652 | auto aliasOp = |
| 3653 | mlir::LLVM::AliasOp::create(builder&: rewriter, location: loc, type: llvmType, linkage, name: symbol, |
| 3654 | dsoLocal: isDsoLocal, tls_mode: threadLocalMode, attrs: attributes); |
| 3655 | { |
| 3656 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 3657 | rewriter.createBlock(parent: &aliasOp.getInitializerRegion()); |
| 3658 | mlir::Type ptrTy = |
| 3659 | mlir::LLVM::LLVMPointerType::get(context: getContext(), addressSpace: addrSpace); |
| 3660 | auto addrOp = |
| 3661 | mlir::LLVM::AddressOfOp::create(builder&: rewriter, location: loc, res: ptrTy, global_name: *aliasee); |
| 3662 | mlir::LLVM::ReturnOp::create(builder&: rewriter, location: loc, arg: addrOp); |
| 3663 | } |
| 3664 | symbolTables.getSymbolTable(op: modOp).remove(op); |
| 3665 | rewriter.replaceOp(op, newOp: aliasOp); |
| 3666 | symbolTables.getSymbolTable(op: modOp).insert(symbol: aliasOp); |
| 3667 | return mlir::success(); |
| 3668 | } |
| 3669 | |
| 3670 | if (init.has_value()) { |
| 3671 | if (mlir::isa<cir::FPAttr, cir::IntAttr, cir::BoolAttr>(Val: init.value())) { |
| 3672 | GlobalInitAttrRewriter initRewriter(llvmType, dataLayout, rewriter); |
| 3673 | init = initRewriter.visit(attr: init.value()); |
| 3674 | // If initRewriter returned a null attribute, init will have a value but |
| 3675 | // the value will be null. If that happens, initRewriter didn't handle the |
| 3676 | // attribute type. It probably needs to be added to |
| 3677 | // GlobalInitAttrRewriter. |
| 3678 | if (!init.value()) { |
| 3679 | op.emitError() << "unsupported initializer '" << init.value() << "'" ; |
| 3680 | return mlir::failure(); |
| 3681 | } |
| 3682 | } else if (auto constArr = |
| 3683 | mlir::dyn_cast<cir::ConstArrayAttr>(Val&: init.value())) { |
| 3684 | // Bulk-emit llvm.mlir.global when lowerConstArrayAttr can build the |
| 3685 | // whole initializer as one aggregate attribute (no insertvalue |
| 3686 | // region). Leaf type must match what lowerConstArrayAttr handles |
| 3687 | // (pointers, integers, bools, floats, and string literals with |
| 3688 | // trailing_zeros). |
| 3689 | if (isBulkLowerableConstArrayBaseElement( |
| 3690 | baseElemTy: getConstArrayBaseElementType(ty: constArr.getType()))) { |
| 3691 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3692 | if (std::optional<mlir::Attribute> bulkInit = |
| 3693 | lowerConstArrayAttr(constArr, symbolTables, converter: typeConverter, |
| 3694 | preConversionTypes, moduleOp: modOp)) { |
| 3695 | mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, builder&: rewriter); |
| 3696 | replaceCirGlobal(symbolTables, op, llvmType, isConst, linkage, symbol, |
| 3697 | init: bulkInit.value(), alignment, addrSpace, isDsoLocal, |
| 3698 | threadLocalMode, comdatAttr, attributes, rewriter); |
| 3699 | return mlir::success(); |
| 3700 | } |
| 3701 | } |
| 3702 | return matchAndRewriteRegionInitializedGlobal(op, init: init.value(), rewriter); |
| 3703 | } else if (auto constRecord = |
| 3704 | mlir::dyn_cast<cir::ConstRecordAttr>(Val&: init.value())) { |
| 3705 | // Bulk-emit llvm.mlir.global when every member of the record can be |
| 3706 | // lowered to a constant attribute. The LLVM dialect global translation |
| 3707 | // turns an ArrayAttr (one element per struct field) into an |
| 3708 | // llvm::ConstantStruct, so the whole initializer becomes a single |
| 3709 | // attribute on the global instead of an insertvalue region. |
| 3710 | mlir::ModuleOp modOp = op->getParentOfType<mlir::ModuleOp>(); |
| 3711 | if (std::optional<mlir::Attribute> bulkInit = |
| 3712 | lowerConstRecordAttr(constRecord, symbolTables, converter: typeConverter, |
| 3713 | preConversionTypes, moduleOp: modOp)) { |
| 3714 | mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, builder&: rewriter); |
| 3715 | replaceCirGlobal(symbolTables, op, llvmType, isConst, linkage, symbol, |
| 3716 | init: bulkInit.value(), alignment, addrSpace, isDsoLocal, |
| 3717 | threadLocalMode, comdatAttr, attributes, rewriter); |
| 3718 | return mlir::success(); |
| 3719 | } |
| 3720 | return matchAndRewriteRegionInitializedGlobal(op, init: init.value(), rewriter); |
| 3721 | } else if (mlir::isa<cir::BlockAddrDiffAttr, cir::BlockAddrInfoAttr, |
| 3722 | cir::ConstVectorAttr, cir::ConstRecordAttr, |
| 3723 | cir::ConstPtrAttr, cir::ConstComplexAttr, |
| 3724 | cir::GlobalOffsetAttr, cir::GlobalViewAttr, |
| 3725 | cir::TypeInfoAttr, cir::UndefAttr, cir::PoisonAttr, |
| 3726 | cir::VTableAttr, cir::ZeroAttr>(Val: init.value())) { |
| 3727 | // TODO(cir): once LLVM's dialect has proper equivalent attributes this |
| 3728 | // should be updated. For now, we use a custom op to initialize globals |
| 3729 | // to the appropriate value. |
| 3730 | return matchAndRewriteRegionInitializedGlobal(op, init: init.value(), rewriter); |
| 3731 | } else { |
| 3732 | // We will only get here if new initializer types are added and this |
| 3733 | // code is not updated to handle them. |
| 3734 | op.emitError() << "unsupported initializer '" << init.value() << "'" ; |
| 3735 | return mlir::failure(); |
| 3736 | } |
| 3737 | } |
| 3738 | |
| 3739 | mlir::SymbolRefAttr comdatAttr = getComdatAttr(op, builder&: rewriter); |
| 3740 | replaceCirGlobal(symbolTables, op, llvmType, isConst, linkage, symbol, |
| 3741 | init: init.value_or(u: mlir::Attribute()), alignment, addrSpace, |
| 3742 | isDsoLocal, threadLocalMode, comdatAttr, attributes, |
| 3743 | rewriter); |
| 3744 | |
| 3745 | return mlir::success(); |
| 3746 | } |
| 3747 | |
| 3748 | static mlir::SymbolRefAttr |
| 3749 | getComdatAttrHelper(mlir::ModuleOp modOp, mlir::OpBuilder &builder, |
| 3750 | StringRef symName, mlir::LLVM::ComdatOp &comdatOp, |
| 3751 | mlir::SymbolTableCollection &symbolTables) { |
| 3752 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 3753 | StringRef comdatName = "__llvm_comdat" ; |
| 3754 | if (!comdatOp) { |
| 3755 | // The GlobalOp and FuncOp lowering patterns each cache their own |
| 3756 | // 'comdatOp', but both now share a single module-level comdat region, so |
| 3757 | // whichever pattern gets here second has to find the existing one rather |
| 3758 | // than create a duplicate symbol. |
| 3759 | comdatOp = modOp.lookupSymbol<mlir::LLVM::ComdatOp>(name: comdatName); |
| 3760 | } |
| 3761 | |
| 3762 | if (!comdatOp) { |
| 3763 | builder.setInsertionPointToStart(modOp.getBody()); |
| 3764 | comdatOp = |
| 3765 | mlir::LLVM::ComdatOp::create(builder, location: modOp.getLoc(), symName: comdatName); |
| 3766 | } |
| 3767 | |
| 3768 | // Cached, and shared by both patterns: a linear scan of the comdat region |
| 3769 | // per symbol is quadratic in the number of comdats. |
| 3770 | mlir::SymbolTable &selectors = symbolTables.getSymbolTable(op: comdatOp); |
| 3771 | if (auto comdatSelector = |
| 3772 | selectors.lookup<mlir::LLVM::ComdatSelectorOp>(name: symName)) { |
| 3773 | return mlir::SymbolRefAttr::get( |
| 3774 | ctx: builder.getContext(), value: comdatName, |
| 3775 | nestedRefs: mlir::FlatSymbolRefAttr::get(value: comdatSelector.getSymNameAttr())); |
| 3776 | } |
| 3777 | |
| 3778 | builder.setInsertionPointToStart(&comdatOp.getBody().back()); |
| 3779 | auto selectorOp = mlir::LLVM::ComdatSelectorOp::create( |
| 3780 | builder, location: comdatOp.getLoc(), sym_name: symName, comdat: mlir::LLVM::comdat::Comdat::Any, |
| 3781 | /*sym_visibility=*/nullptr); |
| 3782 | selectors.insert(symbol: selectorOp); |
| 3783 | return mlir::SymbolRefAttr::get( |
| 3784 | ctx: builder.getContext(), value: comdatName, |
| 3785 | nestedRefs: mlir::FlatSymbolRefAttr::get(value: selectorOp.getSymNameAttr())); |
| 3786 | } |
| 3787 | |
| 3788 | mlir::SymbolRefAttr |
| 3789 | CIRToLLVMGlobalOpLowering::getComdatAttr(cir::GlobalOp &op, |
| 3790 | mlir::OpBuilder &builder) const { |
| 3791 | std::optional<llvm::StringRef> comdat = op.getComdat(); |
| 3792 | if (!comdat) |
| 3793 | return mlir::SymbolRefAttr{}; |
| 3794 | llvm::StringRef comdatKey = comdat->empty() ? op.getSymName() : *comdat; |
| 3795 | return getComdatAttrHelper(modOp: op->getParentOfType<mlir::ModuleOp>(), builder, |
| 3796 | symName: comdatKey, comdatOp, symbolTables); |
| 3797 | } |
| 3798 | |
| 3799 | mlir::SymbolRefAttr |
| 3800 | CIRToLLVMFuncOpLowering::getComdatAttr(cir::FuncOp &op, |
| 3801 | mlir::OpBuilder &builder) const { |
| 3802 | std::optional<llvm::StringRef> comdat = op.getComdat(); |
| 3803 | if (!comdat) |
| 3804 | return mlir::SymbolRefAttr{}; |
| 3805 | llvm::StringRef comdatKey = comdat->empty() ? op.getSymName() : *comdat; |
| 3806 | return getComdatAttrHelper(modOp: op->getParentOfType<mlir::ModuleOp>(), builder, |
| 3807 | symName: comdatKey, comdatOp, symbolTables); |
| 3808 | } |
| 3809 | |
| 3810 | mlir::LogicalResult CIRToLLVMSwitchFlatOpLowering::matchAndRewrite( |
| 3811 | cir::SwitchFlatOp op, OpAdaptor adaptor, |
| 3812 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3813 | |
| 3814 | llvm::SmallVector<mlir::APInt, 8> caseValues; |
| 3815 | for (mlir::Attribute val : op.getCaseValues()) { |
| 3816 | auto intAttr = cast<cir::IntAttr>(Val&: val); |
| 3817 | caseValues.push_back(Elt: intAttr.getValue()); |
| 3818 | } |
| 3819 | |
| 3820 | llvm::SmallVector<mlir::Block *, 8> caseDestinations; |
| 3821 | |
| 3822 | for (mlir::Block *x : op.getCaseDestinations()) |
| 3823 | caseDestinations.push_back(Elt: x); |
| 3824 | |
| 3825 | // Use the converted operands so they match the type-converted successor |
| 3826 | // block arguments. |
| 3827 | llvm::SmallVector<mlir::ValueRange> caseOperands = adaptor.getCaseOperands(); |
| 3828 | |
| 3829 | // Set switch op to branch to the newly created blocks. |
| 3830 | rewriter.setInsertionPoint(op); |
| 3831 | rewriter.replaceOpWithNewOp<mlir::LLVM::SwitchOp>( |
| 3832 | op, args: adaptor.getCondition(), args: op.getDefaultDestination(), |
| 3833 | args: adaptor.getDefaultOperands(), args&: caseValues, args&: caseDestinations, args&: caseOperands); |
| 3834 | return mlir::success(); |
| 3835 | } |
| 3836 | |
| 3837 | static mlir::LLVM::IntegerOverflowFlags nswFlag(bool nsw) { |
| 3838 | return nsw ? mlir::LLVM::IntegerOverflowFlags::nsw |
| 3839 | : mlir::LLVM::IntegerOverflowFlags::none; |
| 3840 | } |
| 3841 | |
| 3842 | template <typename CIROp, typename LLVMIntOp> |
| 3843 | static mlir::LogicalResult |
| 3844 | lowerIncDecOp(CIROp op, typename CIROp::Adaptor adaptor, |
| 3845 | mlir::ConversionPatternRewriter &rewriter) { |
| 3846 | mlir::Type llvmType = adaptor.getInput().getType(); |
| 3847 | mlir::Location loc = op.getLoc(); |
| 3848 | |
| 3849 | auto maybeNSW = nswFlag(op.getNoSignedWrap()); |
| 3850 | mlir::LLVM::ConstantOp one; |
| 3851 | if (mlir::isa<cir::VectorType>(op.getType())) { |
| 3852 | mlir::ShapedType shapedTy = mlir::cast<mlir::ShapedType>(Val&: llvmType); |
| 3853 | mlir::APInt oneAP(shapedTy.getElementTypeBitWidth(), 1); |
| 3854 | auto oneVec = mlir::DenseElementsAttr::get(type: shapedTy, list: {oneAP}); |
| 3855 | one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: llvmType, value: oneVec); |
| 3856 | } else { |
| 3857 | one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: llvmType, value: 1); |
| 3858 | } |
| 3859 | rewriter.replaceOpWithNewOp<LLVMIntOp>(op, adaptor.getInput(), one, maybeNSW); |
| 3860 | return mlir::success(); |
| 3861 | } |
| 3862 | |
| 3863 | mlir::LogicalResult CIRToLLVMIncOpLowering::matchAndRewrite( |
| 3864 | cir::IncOp op, OpAdaptor adaptor, |
| 3865 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3866 | return lowerIncDecOp<cir::IncOp, mlir::LLVM::AddOp>(op, adaptor, rewriter); |
| 3867 | } |
| 3868 | |
| 3869 | mlir::LogicalResult CIRToLLVMDecOpLowering::matchAndRewrite( |
| 3870 | cir::DecOp op, OpAdaptor adaptor, |
| 3871 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3872 | return lowerIncDecOp<cir::DecOp, mlir::LLVM::SubOp>(op, adaptor, rewriter); |
| 3873 | } |
| 3874 | |
| 3875 | mlir::LogicalResult CIRToLLVMMinusOpLowering::matchAndRewrite( |
| 3876 | cir::MinusOp op, OpAdaptor adaptor, |
| 3877 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3878 | bool isVector = mlir::isa<cir::VectorType>(Val: op.getType()); |
| 3879 | mlir::Type llvmType = adaptor.getInput().getType(); |
| 3880 | mlir::Location loc = op.getLoc(); |
| 3881 | |
| 3882 | auto maybeNSW = nswFlag(nsw: op.getNoSignedWrap()); |
| 3883 | mlir::Value zero; |
| 3884 | if (isVector) |
| 3885 | zero = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: llvmType); |
| 3886 | else |
| 3887 | zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: llvmType, value: 0); |
| 3888 | rewriter.replaceOpWithNewOp<mlir::LLVM::SubOp>(op, args&: zero, args: adaptor.getInput(), |
| 3889 | args&: maybeNSW); |
| 3890 | return mlir::success(); |
| 3891 | } |
| 3892 | |
| 3893 | mlir::LogicalResult CIRToLLVMNotOpLowering::matchAndRewrite( |
| 3894 | cir::NotOp op, OpAdaptor adaptor, |
| 3895 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3896 | mlir::Type elementType = elementTypeIfVectorOrMatrix(type: op.getType()); |
| 3897 | bool isVector = mlir::isa<cir::VectorType>(Val: op.getType()); |
| 3898 | mlir::Type llvmType = adaptor.getInput().getType(); |
| 3899 | mlir::Location loc = op.getLoc(); |
| 3900 | |
| 3901 | if (mlir::isa<cir::IntType>(Val: elementType)) { |
| 3902 | mlir::Value minusOne; |
| 3903 | if (isVector) { |
| 3904 | const unsigned eltWidth = |
| 3905 | mlir::cast<cir::IntType>(Val&: elementType).getWidth(); |
| 3906 | mlir::DenseElementsAttr denseVec = |
| 3907 | mlir::DenseElementsAttr::get(type: mlir::cast<mlir::ShapedType>(Val&: llvmType), |
| 3908 | list: {APInt::getAllOnes(numBits: eltWidth)}); |
| 3909 | minusOne = |
| 3910 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: llvmType, value: denseVec); |
| 3911 | } else { |
| 3912 | minusOne = mlir::LLVM::ConstantOp::create( |
| 3913 | builder&: rewriter, location: loc, type: llvmType, |
| 3914 | value: APInt::getAllOnes(numBits: cast<mlir::IntegerType>(Val&: llvmType).getWidth())); |
| 3915 | } |
| 3916 | rewriter.replaceOpWithNewOp<mlir::LLVM::XOrOp>(op, args: adaptor.getInput(), |
| 3917 | args&: minusOne); |
| 3918 | return mlir::success(); |
| 3919 | } |
| 3920 | |
| 3921 | if (mlir::isa<cir::BoolType>(Val: elementType)) { |
| 3922 | mlir::Value one; |
| 3923 | if (isVector) { |
| 3924 | mlir::DenseElementsAttr denseVec = mlir::DenseElementsAttr::get( |
| 3925 | type: mlir::cast<mlir::ShapedType>(Val&: llvmType), list: {APInt(1, 1)}); |
| 3926 | one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: llvmType, value: denseVec); |
| 3927 | } else { |
| 3928 | one = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: llvmType, value: 1); |
| 3929 | } |
| 3930 | rewriter.replaceOpWithNewOp<mlir::LLVM::XOrOp>(op, args: adaptor.getInput(), args&: one); |
| 3931 | return mlir::success(); |
| 3932 | } |
| 3933 | |
| 3934 | return op.emitError() << "Unsupported type for bitwise NOT" ; |
| 3935 | } |
| 3936 | |
| 3937 | static bool isIntTypeUnsigned(mlir::Type type) { |
| 3938 | // TODO: Ideally, we should only need to check cir::IntType here. |
| 3939 | return mlir::isa<cir::IntType>(Val: type) |
| 3940 | ? mlir::cast<cir::IntType>(Val&: type).isUnsigned() |
| 3941 | : mlir::cast<mlir::IntegerType>(Val&: type).isUnsigned(); |
| 3942 | } |
| 3943 | |
| 3944 | //===----------------------------------------------------------------------===// |
| 3945 | // Binary Op Lowering |
| 3946 | //===----------------------------------------------------------------------===// |
| 3947 | |
| 3948 | mlir::LLVM::IntegerOverflowFlags |
| 3949 | convertToLLVMProperty(mlir::MLIRContext *, cir::OverflowFlags flags) { |
| 3950 | mlir::LLVM::IntegerOverflowFlags result = |
| 3951 | mlir::LLVM::IntegerOverflowFlags::none; |
| 3952 | if (bitEnumContainsAny(bits: flags, bit: cir::OverflowFlags::NoSignedWrap)) |
| 3953 | result = result | mlir::LLVM::IntegerOverflowFlags::nsw; |
| 3954 | if (bitEnumContainsAny(bits: flags, bit: cir::OverflowFlags::NoUnsignedWrap)) |
| 3955 | result = result | mlir::LLVM::IntegerOverflowFlags::nuw; |
| 3956 | return result; |
| 3957 | } |
| 3958 | |
| 3959 | /// Lower an integer op to its signed or unsigned LLVM counterpart. |
| 3960 | template <typename UIntOp, typename SIntOp, typename CIROp> |
| 3961 | static mlir::LogicalResult |
| 3962 | lowerIntBinaryOp(CIROp op, mlir::Value lhs, mlir::Value rhs, |
| 3963 | mlir::ConversionPatternRewriter &rewriter) { |
| 3964 | const mlir::Type eltType = elementTypeIfVectorOrMatrix(op.getRhs().getType()); |
| 3965 | assert(cir::isIntOrBoolType(eltType) && |
| 3966 | "integer binary op expects integer operand types" ); |
| 3967 | if (isIntTypeUnsigned(type: eltType)) |
| 3968 | rewriter.replaceOpWithNewOp<UIntOp>(op, lhs, rhs); |
| 3969 | else |
| 3970 | rewriter.replaceOpWithNewOp<SIntOp>(op, lhs, rhs); |
| 3971 | return mlir::success(); |
| 3972 | } |
| 3973 | |
| 3974 | mlir::LogicalResult CIRToLLVMAddSatOpLowering::matchAndRewrite( |
| 3975 | cir::AddSatOp op, OpAdaptor adaptor, |
| 3976 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3977 | return lowerIntBinaryOp<mlir::LLVM::UAddSat, mlir::LLVM::SAddSat>( |
| 3978 | op, lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), rewriter); |
| 3979 | } |
| 3980 | |
| 3981 | mlir::LogicalResult CIRToLLVMSubSatOpLowering::matchAndRewrite( |
| 3982 | cir::SubSatOp op, OpAdaptor adaptor, |
| 3983 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3984 | return lowerIntBinaryOp<mlir::LLVM::USubSat, mlir::LLVM::SSubSat>( |
| 3985 | op, lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), rewriter); |
| 3986 | } |
| 3987 | |
| 3988 | mlir::LogicalResult CIRToLLVMDivOpLowering::matchAndRewrite( |
| 3989 | cir::DivOp op, OpAdaptor adaptor, |
| 3990 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3991 | return lowerIntBinaryOp<mlir::LLVM::UDivOp, mlir::LLVM::SDivOp>( |
| 3992 | op, lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), rewriter); |
| 3993 | } |
| 3994 | |
| 3995 | mlir::LogicalResult CIRToLLVMRemOpLowering::matchAndRewrite( |
| 3996 | cir::RemOp op, OpAdaptor adaptor, |
| 3997 | mlir::ConversionPatternRewriter &rewriter) const { |
| 3998 | return lowerIntBinaryOp<mlir::LLVM::URemOp, mlir::LLVM::SRemOp>( |
| 3999 | op, lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), rewriter); |
| 4000 | } |
| 4001 | |
| 4002 | template <typename CIROp, typename UIntOp, typename SIntOp> |
| 4003 | static mlir::LogicalResult |
| 4004 | lowerMinMaxOp(CIROp op, typename CIROp::Adaptor adaptor, |
| 4005 | mlir::ConversionPatternRewriter &rewriter) { |
| 4006 | const mlir::Value lhs = adaptor.getLhs(); |
| 4007 | const mlir::Value rhs = adaptor.getRhs(); |
| 4008 | if (isIntTypeUnsigned(elementTypeIfVectorOrMatrix(op.getRhs().getType()))) |
| 4009 | rewriter.replaceOpWithNewOp<UIntOp>(op, lhs, rhs); |
| 4010 | else |
| 4011 | rewriter.replaceOpWithNewOp<SIntOp>(op, lhs, rhs); |
| 4012 | return mlir::success(); |
| 4013 | } |
| 4014 | |
| 4015 | mlir::LogicalResult CIRToLLVMMaxOpLowering::matchAndRewrite( |
| 4016 | cir::MaxOp op, OpAdaptor adaptor, |
| 4017 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4018 | return lowerMinMaxOp<cir::MaxOp, mlir::LLVM::UMaxOp, mlir::LLVM::SMaxOp>( |
| 4019 | op, adaptor, rewriter); |
| 4020 | } |
| 4021 | |
| 4022 | mlir::LogicalResult CIRToLLVMMinOpLowering::matchAndRewrite( |
| 4023 | cir::MinOp op, OpAdaptor adaptor, |
| 4024 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4025 | return lowerMinMaxOp<cir::MinOp, mlir::LLVM::UMinOp, mlir::LLVM::SMinOp>( |
| 4026 | op, adaptor, rewriter); |
| 4027 | } |
| 4028 | |
| 4029 | /// Convert from a CIR comparison kind to an LLVM IR integral comparison kind. |
| 4030 | static mlir::LLVM::ICmpPredicate |
| 4031 | convertCmpKindToICmpPredicate(cir::CmpOpKind kind, bool isSigned) { |
| 4032 | using CIR = cir::CmpOpKind; |
| 4033 | using LLVMICmp = mlir::LLVM::ICmpPredicate; |
| 4034 | switch (kind) { |
| 4035 | case CIR::eq: |
| 4036 | return LLVMICmp::eq; |
| 4037 | case CIR::ne: |
| 4038 | return LLVMICmp::ne; |
| 4039 | case CIR::lt: |
| 4040 | return (isSigned ? LLVMICmp::slt : LLVMICmp::ult); |
| 4041 | case CIR::le: |
| 4042 | return (isSigned ? LLVMICmp::sle : LLVMICmp::ule); |
| 4043 | case CIR::gt: |
| 4044 | return (isSigned ? LLVMICmp::sgt : LLVMICmp::ugt); |
| 4045 | case CIR::ge: |
| 4046 | return (isSigned ? LLVMICmp::sge : LLVMICmp::uge); |
| 4047 | case CIR::one: |
| 4048 | case CIR::uno: |
| 4049 | llvm_unreachable("FP-only comparison used with integer type" ); |
| 4050 | } |
| 4051 | llvm_unreachable("Unknown CmpOpKind" ); |
| 4052 | } |
| 4053 | |
| 4054 | /// Convert from a CIR comparison kind to an LLVM IR floating-point comparison |
| 4055 | /// kind. |
| 4056 | static mlir::LLVM::FCmpPredicate |
| 4057 | convertCmpKindToFCmpPredicate(cir::CmpOpKind kind) { |
| 4058 | using CIR = cir::CmpOpKind; |
| 4059 | using LLVMFCmp = mlir::LLVM::FCmpPredicate; |
| 4060 | switch (kind) { |
| 4061 | case CIR::eq: |
| 4062 | return LLVMFCmp::oeq; |
| 4063 | case CIR::ne: |
| 4064 | return LLVMFCmp::une; |
| 4065 | case CIR::lt: |
| 4066 | return LLVMFCmp::olt; |
| 4067 | case CIR::le: |
| 4068 | return LLVMFCmp::ole; |
| 4069 | case CIR::gt: |
| 4070 | return LLVMFCmp::ogt; |
| 4071 | case CIR::ge: |
| 4072 | return LLVMFCmp::oge; |
| 4073 | case CIR::one: |
| 4074 | return LLVMFCmp::one; |
| 4075 | case CIR::uno: |
| 4076 | return LLVMFCmp::uno; |
| 4077 | } |
| 4078 | llvm_unreachable("Unknown CmpOpKind" ); |
| 4079 | } |
| 4080 | |
| 4081 | /// Convert from a CIR fp comparison kind to an LLVM IR floating-point |
| 4082 | /// comparison kind. |
| 4083 | static mlir::LLVM::FCmpPredicate |
| 4084 | convertCmpKindToFCmpPredicate(cir::FCmpOpKind kind) { |
| 4085 | using CIR = cir::FCmpOpKind; |
| 4086 | using LLVMFCmp = mlir::LLVM::FCmpPredicate; |
| 4087 | switch (kind) { |
| 4088 | case CIR::_false: |
| 4089 | return LLVMFCmp::_false; |
| 4090 | case CIR::oeq: |
| 4091 | return LLVMFCmp::oeq; |
| 4092 | case CIR::ogt: |
| 4093 | return LLVMFCmp::ogt; |
| 4094 | case CIR::oge: |
| 4095 | return LLVMFCmp::oge; |
| 4096 | case CIR::olt: |
| 4097 | return LLVMFCmp::olt; |
| 4098 | case CIR::ole: |
| 4099 | return LLVMFCmp::ole; |
| 4100 | case CIR::one: |
| 4101 | return LLVMFCmp::one; |
| 4102 | case CIR::ord: |
| 4103 | return LLVMFCmp::ord; |
| 4104 | case CIR::ueq: |
| 4105 | return LLVMFCmp::ueq; |
| 4106 | case CIR::ugt: |
| 4107 | return LLVMFCmp::ugt; |
| 4108 | case CIR::uge: |
| 4109 | return LLVMFCmp::uge; |
| 4110 | case CIR::ult: |
| 4111 | return LLVMFCmp::ult; |
| 4112 | case CIR::ule: |
| 4113 | return LLVMFCmp::ule; |
| 4114 | case CIR::une: |
| 4115 | return LLVMFCmp::une; |
| 4116 | case CIR::uno: |
| 4117 | return LLVMFCmp::uno; |
| 4118 | case CIR::_true: |
| 4119 | return LLVMFCmp::_true; |
| 4120 | } |
| 4121 | llvm_unreachable("Unknown FCmpOpKind" ); |
| 4122 | } |
| 4123 | |
| 4124 | static llvm::StringRef |
| 4125 | convertCmpKindToConstrainedFCmpPredicate(cir::CmpOpKind kind) { |
| 4126 | using CIR = cir::CmpOpKind; |
| 4127 | switch (kind) { |
| 4128 | case CIR::eq: |
| 4129 | return "oeq" ; |
| 4130 | case CIR::ne: |
| 4131 | return "une" ; |
| 4132 | case CIR::lt: |
| 4133 | return "olt" ; |
| 4134 | case CIR::le: |
| 4135 | return "ole" ; |
| 4136 | case CIR::gt: |
| 4137 | return "ogt" ; |
| 4138 | case CIR::ge: |
| 4139 | return "oge" ; |
| 4140 | case CIR::one: |
| 4141 | return "one" ; |
| 4142 | case CIR::uno: |
| 4143 | return "uno" ; |
| 4144 | } |
| 4145 | llvm_unreachable("Unknown CmpOpKind" ); |
| 4146 | } |
| 4147 | |
| 4148 | static bool isSignalingConstrainedFCmp(cir::CmpOpKind kind) { |
| 4149 | using CIR = cir::CmpOpKind; |
| 4150 | switch (kind) { |
| 4151 | case CIR::lt: |
| 4152 | case CIR::le: |
| 4153 | case CIR::gt: |
| 4154 | case CIR::ge: |
| 4155 | return true; |
| 4156 | case CIR::eq: |
| 4157 | case CIR::ne: |
| 4158 | case CIR::one: |
| 4159 | case CIR::uno: |
| 4160 | return false; |
| 4161 | } |
| 4162 | llvm_unreachable("Unknown CmpOpKind" ); |
| 4163 | } |
| 4164 | |
| 4165 | static llvm::StringRef |
| 4166 | convertCmpKindToConstrainedFCmpPredicate(cir::FCmpOpKind kind) { |
| 4167 | using CIR = cir::FCmpOpKind; |
| 4168 | switch (kind) { |
| 4169 | case CIR::_false: |
| 4170 | return "false" ; |
| 4171 | case CIR::oeq: |
| 4172 | return "oeq" ; |
| 4173 | case CIR::ogt: |
| 4174 | return "ogt" ; |
| 4175 | case CIR::oge: |
| 4176 | return "oge" ; |
| 4177 | case CIR::olt: |
| 4178 | return "olt" ; |
| 4179 | case CIR::ole: |
| 4180 | return "ole" ; |
| 4181 | case CIR::one: |
| 4182 | return "one" ; |
| 4183 | case CIR::ord: |
| 4184 | return "ord" ; |
| 4185 | case CIR::ueq: |
| 4186 | return "ueq" ; |
| 4187 | case CIR::ugt: |
| 4188 | return "ugt" ; |
| 4189 | case CIR::uge: |
| 4190 | return "uge" ; |
| 4191 | case CIR::ult: |
| 4192 | return "ult" ; |
| 4193 | case CIR::ule: |
| 4194 | return "ule" ; |
| 4195 | case CIR::une: |
| 4196 | return "une" ; |
| 4197 | case CIR::uno: |
| 4198 | return "uno" ; |
| 4199 | case CIR::_true: |
| 4200 | return "true" ; |
| 4201 | } |
| 4202 | llvm_unreachable("Unknown FCmpOpKind" ); |
| 4203 | } |
| 4204 | |
| 4205 | template <typename OpKindTy> |
| 4206 | static mlir::LLVM::CallIntrinsicOp |
| 4207 | createConstrainedFCmpCall(mlir::ConversionPatternRewriter &rewriter, |
| 4208 | mlir::Location loc, mlir::Value lhs, mlir::Value rhs, |
| 4209 | OpKindTy kind, bool isSignaling, cir::FenvAttr fenv, |
| 4210 | mlir::Type llvmResTy) { |
| 4211 | llvm::SmallVector<mlir::Value, 4> callOperands = { |
| 4212 | lhs, rhs, |
| 4213 | createFenvMetadataValue(rewriter, loc, |
| 4214 | convertCmpKindToConstrainedFCmpPredicate(kind)), |
| 4215 | createFenvMetadataValue(rewriter, loc, |
| 4216 | str: getConstrainedExceptMetadata(fenv))}; |
| 4217 | llvm::StringRef intrinsicName = isSignaling |
| 4218 | ? "llvm.experimental.constrained.fcmps" |
| 4219 | : "llvm.experimental.constrained.fcmp" ; |
| 4220 | return createCallLLVMIntrinsicOp(rewriter, loc, intrinsicName, resultTy: llvmResTy, |
| 4221 | operands: callOperands); |
| 4222 | } |
| 4223 | |
| 4224 | mlir::LogicalResult CIRToLLVMCmpOpLowering::matchAndRewrite( |
| 4225 | cir::CmpOp cmpOp, OpAdaptor adaptor, |
| 4226 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4227 | mlir::Type type = cmpOp.getLhs().getType(); |
| 4228 | |
| 4229 | if (mlir::isa<cir::IntType, mlir::IntegerType>(Val: type)) { |
| 4230 | bool isSigned = mlir::isa<cir::IntType>(Val: type) |
| 4231 | ? mlir::cast<cir::IntType>(Val&: type).isSigned() |
| 4232 | : mlir::cast<mlir::IntegerType>(Val&: type).isSigned(); |
| 4233 | mlir::LLVM::ICmpPredicate kind = |
| 4234 | convertCmpKindToICmpPredicate(kind: cmpOp.getKind(), isSigned); |
| 4235 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
| 4236 | op: cmpOp, args&: kind, args: adaptor.getLhs(), args: adaptor.getRhs()); |
| 4237 | return mlir::success(); |
| 4238 | } |
| 4239 | |
| 4240 | if (mlir::isa<cir::BoolType, cir::PointerType, cir::VPtrType>(Val: type)) { |
| 4241 | // Booleans, including enums with a boolean underlying type, compare as |
| 4242 | // unsigned integers, as do pointers and !cir.vptr, which lowers to a |
| 4243 | // pointer. |
| 4244 | mlir::LLVM::ICmpPredicate kind = |
| 4245 | convertCmpKindToICmpPredicate(kind: cmpOp.getKind(), /*isSigned=*/false); |
| 4246 | rewriter.replaceOpWithNewOp<mlir::LLVM::ICmpOp>( |
| 4247 | op: cmpOp, args&: kind, args: adaptor.getLhs(), args: adaptor.getRhs()); |
| 4248 | return mlir::success(); |
| 4249 | } |
| 4250 | |
| 4251 | if (mlir::isa<cir::FPTypeInterface>(Val: type)) { |
| 4252 | mlir::Type llvmResTy = getTypeConverter()->convertType(t: cmpOp.getType()); |
| 4253 | if (cir::FenvAttr fenv = cmpOp.getFenvAttr()) { |
| 4254 | mlir::LLVM::CallIntrinsicOp call = createConstrainedFCmpCall( |
| 4255 | rewriter, loc: cmpOp.getLoc(), lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), |
| 4256 | kind: cmpOp.getKind(), isSignaling: isSignalingConstrainedFCmp(kind: cmpOp.getKind()), fenv, |
| 4257 | llvmResTy); |
| 4258 | rewriter.replaceOp(op: cmpOp, newValues: call.getResult(i: 0)); |
| 4259 | return mlir::success(); |
| 4260 | } |
| 4261 | mlir::LLVM::FCmpPredicate kind = |
| 4262 | convertCmpKindToFCmpPredicate(kind: cmpOp.getKind()); |
| 4263 | rewriter.replaceOpWithNewOp<mlir::LLVM::FCmpOp>( |
| 4264 | op: cmpOp, args&: kind, args: adaptor.getLhs(), args: adaptor.getRhs()); |
| 4265 | return mlir::success(); |
| 4266 | } |
| 4267 | |
| 4268 | if (mlir::isa<cir::ComplexType>(Val: type)) { |
| 4269 | mlir::Value lhs = adaptor.getLhs(); |
| 4270 | mlir::Value rhs = adaptor.getRhs(); |
| 4271 | mlir::Location loc = cmpOp.getLoc(); |
| 4272 | |
| 4273 | auto complexType = mlir::cast<cir::ComplexType>(Val: cmpOp.getLhs().getType()); |
| 4274 | mlir::Type complexElemTy = |
| 4275 | getTypeConverter()->convertType(t: complexType.getElementType()); |
| 4276 | |
| 4277 | auto lhsReal = mlir::LLVM::ExtractValueOp::create( |
| 4278 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{0})); |
| 4279 | auto lhsImag = mlir::LLVM::ExtractValueOp::create( |
| 4280 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{1})); |
| 4281 | auto rhsReal = mlir::LLVM::ExtractValueOp::create( |
| 4282 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{0})); |
| 4283 | auto rhsImag = mlir::LLVM::ExtractValueOp::create( |
| 4284 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{1})); |
| 4285 | |
| 4286 | if (cmpOp.getKind() == cir::CmpOpKind::eq) { |
| 4287 | if (complexElemTy.isInteger()) { |
| 4288 | auto realCmp = mlir::LLVM::ICmpOp::create( |
| 4289 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::eq, lhs: lhsReal, rhs: rhsReal); |
| 4290 | auto imagCmp = mlir::LLVM::ICmpOp::create( |
| 4291 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::eq, lhs: lhsImag, rhs: rhsImag); |
| 4292 | rewriter.replaceOpWithNewOp<mlir::LLVM::AndOp>(op: cmpOp, args&: realCmp, args&: imagCmp); |
| 4293 | return mlir::success(); |
| 4294 | } |
| 4295 | |
| 4296 | auto realCmp = mlir::LLVM::FCmpOp::create( |
| 4297 | builder&: rewriter, location: loc, predicate: mlir::LLVM::FCmpPredicate::oeq, lhs: lhsReal, rhs: rhsReal); |
| 4298 | auto imagCmp = mlir::LLVM::FCmpOp::create( |
| 4299 | builder&: rewriter, location: loc, predicate: mlir::LLVM::FCmpPredicate::oeq, lhs: lhsImag, rhs: rhsImag); |
| 4300 | rewriter.replaceOpWithNewOp<mlir::LLVM::AndOp>(op: cmpOp, args&: realCmp, args&: imagCmp); |
| 4301 | return mlir::success(); |
| 4302 | } |
| 4303 | |
| 4304 | if (cmpOp.getKind() == cir::CmpOpKind::ne) { |
| 4305 | if (complexElemTy.isInteger()) { |
| 4306 | auto realCmp = mlir::LLVM::ICmpOp::create( |
| 4307 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::ne, lhs: lhsReal, rhs: rhsReal); |
| 4308 | auto imagCmp = mlir::LLVM::ICmpOp::create( |
| 4309 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::ne, lhs: lhsImag, rhs: rhsImag); |
| 4310 | rewriter.replaceOpWithNewOp<mlir::LLVM::OrOp>(op: cmpOp, args&: realCmp, args&: imagCmp); |
| 4311 | return mlir::success(); |
| 4312 | } |
| 4313 | |
| 4314 | auto realCmp = mlir::LLVM::FCmpOp::create( |
| 4315 | builder&: rewriter, location: loc, predicate: mlir::LLVM::FCmpPredicate::une, lhs: lhsReal, rhs: rhsReal); |
| 4316 | auto imagCmp = mlir::LLVM::FCmpOp::create( |
| 4317 | builder&: rewriter, location: loc, predicate: mlir::LLVM::FCmpPredicate::une, lhs: lhsImag, rhs: rhsImag); |
| 4318 | rewriter.replaceOpWithNewOp<mlir::LLVM::OrOp>(op: cmpOp, args&: realCmp, args&: imagCmp); |
| 4319 | return mlir::success(); |
| 4320 | } |
| 4321 | } |
| 4322 | |
| 4323 | return cmpOp.emitError() << "unsupported type for CmpOp: " << type; |
| 4324 | } |
| 4325 | |
| 4326 | /// Shared lowering logic for checked binary arithmetic overflow operations. |
| 4327 | /// The \p opStr parameter specifies the arithmetic operation name used in the |
| 4328 | /// LLVM intrinsic (e.g., "add", "sub", "mul"). |
| 4329 | template <typename OpTy> |
| 4330 | static mlir::LogicalResult |
| 4331 | lowerBinOpOverflow(OpTy op, typename OpTy::Adaptor adaptor, |
| 4332 | mlir::ConversionPatternRewriter &rewriter, |
| 4333 | const mlir::TypeConverter *typeConverter, |
| 4334 | llvm::StringRef opStr) { |
| 4335 | mlir::Location loc = op.getLoc(); |
| 4336 | cir::IntType operandTy = op.getLhs().getType(); |
| 4337 | // The result type may be a `cir.bool`, which behaves as a 1-bit unsigned |
| 4338 | // integer for the purposes of the checked arithmetic. |
| 4339 | mlir::Type resultTy = op.getResult().getType(); |
| 4340 | auto resultIntTy = mlir::dyn_cast<cir::IntType>(Val&: resultTy); |
| 4341 | unsigned resultWidth = resultIntTy ? resultIntTy.getWidth() : 1; |
| 4342 | bool resultSigned = resultIntTy && resultIntTy.getIsSigned(); |
| 4343 | |
| 4344 | bool sign = operandTy.getIsSigned() || resultSigned; |
| 4345 | unsigned width = |
| 4346 | std::max(a: operandTy.getWidth() + (sign && operandTy.isUnsigned()), |
| 4347 | b: resultWidth + (sign && !resultSigned)); |
| 4348 | |
| 4349 | mlir::IntegerType encompassedLLVMTy = rewriter.getIntegerType(width); |
| 4350 | |
| 4351 | mlir::Value lhs = adaptor.getLhs(); |
| 4352 | mlir::Value rhs = adaptor.getRhs(); |
| 4353 | if (operandTy.getWidth() < width) { |
| 4354 | if (operandTy.isSigned()) { |
| 4355 | lhs = mlir::LLVM::SExtOp::create(builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: lhs); |
| 4356 | rhs = mlir::LLVM::SExtOp::create(builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: rhs); |
| 4357 | } else { |
| 4358 | lhs = mlir::LLVM::ZExtOp::create(builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: lhs); |
| 4359 | rhs = mlir::LLVM::ZExtOp::create(builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: rhs); |
| 4360 | } |
| 4361 | } |
| 4362 | |
| 4363 | // The intrinsic name is `@llvm.{s|u}{op}.with.overflow.i{width}` |
| 4364 | std::string intrinName = ("llvm." + llvm::Twine(sign ? 's' : 'u') + opStr + |
| 4365 | ".with.overflow.i" + llvm::Twine(width)) |
| 4366 | .str(); |
| 4367 | auto intrinNameAttr = mlir::StringAttr::get(op.getContext(), intrinName); |
| 4368 | |
| 4369 | mlir::IntegerType overflowLLVMTy = rewriter.getI1Type(); |
| 4370 | auto intrinRetTy = mlir::LLVM::LLVMStructType::getLiteral( |
| 4371 | context: rewriter.getContext(), types: {encompassedLLVMTy, overflowLLVMTy}); |
| 4372 | |
| 4373 | auto callLLVMIntrinOp = mlir::LLVM::CallIntrinsicOp::create( |
| 4374 | rewriter, loc, intrinRetTy, intrinNameAttr, mlir::ValueRange{lhs, rhs}); |
| 4375 | mlir::Value intrinRet = callLLVMIntrinOp.getResult(0); |
| 4376 | |
| 4377 | mlir::Value result = mlir::LLVM::ExtractValueOp::create( |
| 4378 | builder&: rewriter, location: loc, container: intrinRet, position: ArrayRef<int64_t>{0}) |
| 4379 | .getResult(); |
| 4380 | mlir::Value overflow = mlir::LLVM::ExtractValueOp::create( |
| 4381 | builder&: rewriter, location: loc, container: intrinRet, position: ArrayRef<int64_t>{1}) |
| 4382 | .getResult(); |
| 4383 | |
| 4384 | if (resultWidth < width) { |
| 4385 | mlir::Type resultLLVMTy = typeConverter->convertType(t: resultTy); |
| 4386 | auto truncResult = |
| 4387 | mlir::LLVM::TruncOp::create(builder&: rewriter, location: loc, res: resultLLVMTy, arg: result); |
| 4388 | |
| 4389 | // Extend the truncated result back to the encompassing type to check for |
| 4390 | // any overflows during the truncation. |
| 4391 | mlir::Value truncResultExt; |
| 4392 | if (resultSigned) |
| 4393 | truncResultExt = mlir::LLVM::SExtOp::create( |
| 4394 | builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: truncResult); |
| 4395 | else |
| 4396 | truncResultExt = mlir::LLVM::ZExtOp::create( |
| 4397 | builder&: rewriter, location: loc, res: encompassedLLVMTy, arg: truncResult); |
| 4398 | auto truncOverflow = mlir::LLVM::ICmpOp::create( |
| 4399 | builder&: rewriter, location: loc, predicate: mlir::LLVM::ICmpPredicate::ne, lhs: truncResultExt, rhs: result); |
| 4400 | |
| 4401 | result = truncResult; |
| 4402 | overflow = mlir::LLVM::OrOp::create(builder&: rewriter, location: loc, lhs: overflow, rhs: truncOverflow); |
| 4403 | } |
| 4404 | |
| 4405 | mlir::Type boolLLVMTy = |
| 4406 | typeConverter->convertType(op.getOverflow().getType()); |
| 4407 | if (boolLLVMTy != rewriter.getI1Type()) |
| 4408 | overflow = mlir::LLVM::ZExtOp::create(builder&: rewriter, location: loc, res: boolLLVMTy, arg: overflow); |
| 4409 | |
| 4410 | rewriter.replaceOp(op, mlir::ValueRange{result, overflow}); |
| 4411 | |
| 4412 | return mlir::success(); |
| 4413 | } |
| 4414 | |
| 4415 | mlir::LogicalResult CIRToLLVMAddOverflowOpLowering::matchAndRewrite( |
| 4416 | cir::AddOverflowOp op, OpAdaptor adaptor, |
| 4417 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4418 | return lowerBinOpOverflow(op, adaptor, rewriter, typeConverter: getTypeConverter(), opStr: "add" ); |
| 4419 | } |
| 4420 | |
| 4421 | mlir::LogicalResult CIRToLLVMSubOverflowOpLowering::matchAndRewrite( |
| 4422 | cir::SubOverflowOp op, OpAdaptor adaptor, |
| 4423 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4424 | return lowerBinOpOverflow(op, adaptor, rewriter, typeConverter: getTypeConverter(), opStr: "sub" ); |
| 4425 | } |
| 4426 | |
| 4427 | mlir::LogicalResult CIRToLLVMMulOverflowOpLowering::matchAndRewrite( |
| 4428 | cir::MulOverflowOp op, OpAdaptor adaptor, |
| 4429 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4430 | return lowerBinOpOverflow(op, adaptor, rewriter, typeConverter: getTypeConverter(), opStr: "mul" ); |
| 4431 | } |
| 4432 | |
| 4433 | mlir::LogicalResult CIRToLLVMFrexpOpLowering::matchAndRewrite( |
| 4434 | cir::FrexpOp op, OpAdaptor adaptor, |
| 4435 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4436 | mlir::Location loc = op.getLoc(); |
| 4437 | mlir::Type fpLLVMTy = |
| 4438 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 4439 | mlir::Type intLLVMTy = getTypeConverter()->convertType(t: op.getExp().getType()); |
| 4440 | |
| 4441 | auto structTy = mlir::LLVM::LLVMStructType::getLiteral(context: rewriter.getContext(), |
| 4442 | types: {fpLLVMTy, intLLVMTy}); |
| 4443 | |
| 4444 | auto callOp = createCallLLVMIntrinsicOp(rewriter, loc, intrinsicName: "llvm.frexp" , resultTy: structTy, |
| 4445 | operands: adaptor.getSrc()); |
| 4446 | mlir::Value result = callOp.getResult(i: 0); |
| 4447 | |
| 4448 | mlir::Value mantissa = |
| 4449 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: result, position: 0); |
| 4450 | mlir::Value exponent = |
| 4451 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: result, position: 1); |
| 4452 | rewriter.replaceOp(op, newValues: mlir::ValueRange{mantissa, exponent}); |
| 4453 | return mlir::success(); |
| 4454 | } |
| 4455 | |
| 4456 | mlir::LogicalResult CIRToLLVMModfOpLowering::matchAndRewrite( |
| 4457 | cir::ModfOp op, OpAdaptor adaptor, |
| 4458 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4459 | mlir::Location loc = op.getLoc(); |
| 4460 | mlir::Type fpLLVMTy = |
| 4461 | getTypeConverter()->convertType(t: op.getFractional().getType()); |
| 4462 | |
| 4463 | auto structTy = mlir::LLVM::LLVMStructType::getLiteral(context: rewriter.getContext(), |
| 4464 | types: {fpLLVMTy, fpLLVMTy}); |
| 4465 | |
| 4466 | auto callOp = createCallLLVMIntrinsicOp(rewriter, loc, intrinsicName: "llvm.modf" , resultTy: structTy, |
| 4467 | operands: adaptor.getSrc()); |
| 4468 | mlir::Value result = callOp.getResult(i: 0); |
| 4469 | |
| 4470 | mlir::Value fractional = |
| 4471 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: result, position: 0); |
| 4472 | mlir::Value integral = |
| 4473 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: result, position: 1); |
| 4474 | rewriter.replaceOp(op, newValues: mlir::ValueRange{fractional, integral}); |
| 4475 | return mlir::success(); |
| 4476 | } |
| 4477 | |
| 4478 | mlir::LogicalResult CIRToLLVMShiftOpLowering::matchAndRewrite( |
| 4479 | cir::ShiftOp op, OpAdaptor adaptor, |
| 4480 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4481 | assert((op.getValue().getType() == op.getType()) && |
| 4482 | "inconsistent operands' types NYI" ); |
| 4483 | |
| 4484 | const mlir::Type llvmTy = getTypeConverter()->convertType(t: op.getType()); |
| 4485 | mlir::Value amt = adaptor.getAmount(); |
| 4486 | mlir::Value val = adaptor.getValue(); |
| 4487 | |
| 4488 | auto cirAmtTy = mlir::dyn_cast<cir::IntType>(Val: op.getAmount().getType()); |
| 4489 | bool isUnsigned; |
| 4490 | if (cirAmtTy) { |
| 4491 | auto cirValTy = mlir::cast<cir::IntType>(Val: op.getValue().getType()); |
| 4492 | isUnsigned = cirValTy.isUnsigned(); |
| 4493 | |
| 4494 | // Ensure shift amount is the same type as the value. Some undefined |
| 4495 | // behavior might occur in the casts below as per [C99 6.5.7.3]. |
| 4496 | // Vector type shift amount needs no cast as type consistency is expected to |
| 4497 | // be already be enforced at CIRGen. |
| 4498 | if (cirAmtTy) |
| 4499 | amt = getLLVMIntCast(rewriter, llvmSrc: amt, llvmDstIntTy: llvmTy, isUnsigned: true, cirSrcWidth: cirAmtTy.getWidth(), |
| 4500 | cirDstIntWidth: cirValTy.getWidth()); |
| 4501 | } else { |
| 4502 | auto cirValVTy = mlir::cast<cir::VectorType>(Val: op.getValue().getType()); |
| 4503 | isUnsigned = |
| 4504 | mlir::cast<cir::IntType>(Val: cirValVTy.getElementType()).isUnsigned(); |
| 4505 | } |
| 4506 | |
| 4507 | // Lower to the proper LLVM shift operation. |
| 4508 | if (op.getIsShiftleft()) { |
| 4509 | rewriter.replaceOpWithNewOp<mlir::LLVM::ShlOp>(op, args: llvmTy, args&: val, args&: amt); |
| 4510 | return mlir::success(); |
| 4511 | } |
| 4512 | |
| 4513 | if (isUnsigned) |
| 4514 | rewriter.replaceOpWithNewOp<mlir::LLVM::LShrOp>(op, args: llvmTy, args&: val, args&: amt); |
| 4515 | else |
| 4516 | rewriter.replaceOpWithNewOp<mlir::LLVM::AShrOp>(op, args: llvmTy, args&: val, args&: amt); |
| 4517 | return mlir::success(); |
| 4518 | } |
| 4519 | |
| 4520 | mlir::LogicalResult CIRToLLVMSelectOpLowering::matchAndRewrite( |
| 4521 | cir::SelectOp op, OpAdaptor adaptor, |
| 4522 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4523 | // Two special cases in the LLVMIR codegen of select op: |
| 4524 | // - select %0, %1, false => and %0, %1 |
| 4525 | // - select %0, true, %1 => or %0, %1 |
| 4526 | // Match the (possibly already converted) adaptor operands: without pattern |
| 4527 | // rollback the producer may be an LLVM constant by the time this runs. |
| 4528 | if (mlir::isa<cir::BoolType>(Val: op.getType())) { |
| 4529 | if (mlir::matchPattern(value: adaptor.getFalseValue(), pattern: mlir::m_Zero())) { |
| 4530 | // select %0, %1, false => and %0, %1 |
| 4531 | rewriter.replaceOpWithNewOp<mlir::LLVM::AndOp>(op, args: adaptor.getCondition(), |
| 4532 | args: adaptor.getTrueValue()); |
| 4533 | return mlir::success(); |
| 4534 | } |
| 4535 | if (mlir::matchPattern(value: adaptor.getTrueValue(), pattern: mlir::m_One())) { |
| 4536 | // select %0, true, %1 => or %0, %1 |
| 4537 | rewriter.replaceOpWithNewOp<mlir::LLVM::OrOp>(op, args: adaptor.getCondition(), |
| 4538 | args: adaptor.getFalseValue()); |
| 4539 | return mlir::success(); |
| 4540 | } |
| 4541 | } |
| 4542 | |
| 4543 | mlir::Value llvmCondition = adaptor.getCondition(); |
| 4544 | rewriter.replaceOpWithNewOp<mlir::LLVM::SelectOp>( |
| 4545 | op, args&: llvmCondition, args: adaptor.getTrueValue(), args: adaptor.getFalseValue()); |
| 4546 | |
| 4547 | return mlir::success(); |
| 4548 | } |
| 4549 | |
| 4550 | static void prepareTypeConverter(mlir::LLVMTypeConverter &converter, |
| 4551 | mlir::DataLayout &dataLayout) { |
| 4552 | converter.addConversion(callback: [&](cir::PointerType type) -> mlir::Type { |
| 4553 | mlir::ptr::MemorySpaceAttrInterface addrSpaceAttr = type.getAddrSpace(); |
| 4554 | unsigned numericAS = 0; |
| 4555 | |
| 4556 | if (mlir::isa_and_present<cir::LangAddressSpaceAttr>(Val: addrSpaceAttr)) |
| 4557 | return {}; |
| 4558 | |
| 4559 | if (auto targetAsAttr = |
| 4560 | mlir::dyn_cast_if_present<cir::TargetAddressSpaceAttr>( |
| 4561 | Val&: addrSpaceAttr)) |
| 4562 | numericAS = targetAsAttr.getValue(); |
| 4563 | return mlir::LLVM::LLVMPointerType::get(context: type.getContext(), addressSpace: numericAS); |
| 4564 | }); |
| 4565 | converter.addConversion(callback: [&](cir::VPtrType type) -> mlir::Type { |
| 4566 | assert(!cir::MissingFeatures::addressSpace()); |
| 4567 | return mlir::LLVM::LLVMPointerType::get(context: type.getContext()); |
| 4568 | }); |
| 4569 | // On the device side, surface reference is represented as an object handle |
| 4570 | // in 64-bit integer. |
| 4571 | converter.addConversion(callback: [&](cir::CUDADeviceSurfaceType type) -> mlir::Type { |
| 4572 | return mlir::IntegerType::get(context: type.getContext(), width: 64); |
| 4573 | }); |
| 4574 | converter.addConversion(callback: [&](cir::CUDADeviceTextureType type) -> mlir::Type { |
| 4575 | return mlir::IntegerType::get(context: type.getContext(), width: 64); |
| 4576 | }); |
| 4577 | converter.addConversion(callback: [&](cir::ArrayType type) -> mlir::Type { |
| 4578 | mlir::Type ty = |
| 4579 | convertTypeForMemory(converter, dataLayout, type: type.getElementType()); |
| 4580 | // A null element type means an unsupported member (e.g. a _BitInt with |
| 4581 | // byte-array storage); propagate the conversion failure. |
| 4582 | if (!ty) |
| 4583 | return {}; |
| 4584 | return mlir::LLVM::LLVMArrayType::get(elementType: ty, numElements: type.getSize()); |
| 4585 | }); |
| 4586 | converter.addConversion(callback: [&](cir::VectorType type) -> mlir::Type { |
| 4587 | // Vector-of-_BitInt memory layout is not modeled here (the element would |
| 4588 | // stay at its literal width); report a conversion failure rather than emit |
| 4589 | // an unvalidated vector. |
| 4590 | if (auto intTy = mlir::dyn_cast<cir::IntType>(Val: type.getElementType()); |
| 4591 | intTy && intTy.isBitInt()) |
| 4592 | return {}; |
| 4593 | const mlir::Type ty = converter.convertType(t: type.getElementType()); |
| 4594 | if (!ty) |
| 4595 | return {}; |
| 4596 | return mlir::VectorType::get(shape: type.getSize().getKnownMinValue(), elementType: ty, |
| 4597 | scalableDims: {type.isScalable()}); |
| 4598 | }); |
| 4599 | converter.addConversion(callback: [&](cir::MatrixType type) -> mlir::Type { |
| 4600 | const uint64_t size = type.getNumRows() * type.getNumColumns(); |
| 4601 | const mlir::Type elemTy = converter.convertType(t: type.getElementType()); |
| 4602 | return mlir::VectorType::get(shape: size, elementType: elemTy); |
| 4603 | }); |
| 4604 | converter.addConversion(callback: [&](cir::BoolType type) -> mlir::Type { |
| 4605 | return mlir::IntegerType::get(context: type.getContext(), width: 1, |
| 4606 | signedness: mlir::IntegerType::Signless); |
| 4607 | }); |
| 4608 | converter.addConversion(callback: [&](cir::IntType type) -> mlir::Type { |
| 4609 | // LLVM doesn't work with signed types, so we drop the CIR signs here. |
| 4610 | return mlir::IntegerType::get(context: type.getContext(), width: type.getWidth()); |
| 4611 | }); |
| 4612 | converter.addConversion(callback: [&](cir::SingleType type) -> mlir::Type { |
| 4613 | return mlir::Float32Type::get(context: type.getContext()); |
| 4614 | }); |
| 4615 | converter.addConversion(callback: [&](cir::DoubleType type) -> mlir::Type { |
| 4616 | return mlir::Float64Type::get(context: type.getContext()); |
| 4617 | }); |
| 4618 | converter.addConversion(callback: [&](cir::FP80Type type) -> mlir::Type { |
| 4619 | return mlir::Float80Type::get(context: type.getContext()); |
| 4620 | }); |
| 4621 | converter.addConversion(callback: [&](cir::FP128Type type) -> mlir::Type { |
| 4622 | return mlir::Float128Type::get(context: type.getContext()); |
| 4623 | }); |
| 4624 | converter.addConversion(callback: [&](cir::LongDoubleType type) -> mlir::Type { |
| 4625 | return converter.convertType(t: type.getUnderlying()); |
| 4626 | }); |
| 4627 | converter.addConversion(callback: [&](cir::FP16Type type) -> mlir::Type { |
| 4628 | return mlir::Float16Type::get(context: type.getContext()); |
| 4629 | }); |
| 4630 | converter.addConversion(callback: [&](cir::BF16Type type) -> mlir::Type { |
| 4631 | return mlir::BFloat16Type::get(context: type.getContext()); |
| 4632 | }); |
| 4633 | converter.addConversion(callback: [&](cir::ComplexType type) -> mlir::Type { |
| 4634 | // A complex type is lowered to an LLVM struct that contains the real and |
| 4635 | // imaginary part as data fields. |
| 4636 | mlir::Type elementTy = converter.convertType(t: type.getElementType()); |
| 4637 | mlir::Type structFields[2] = {elementTy, elementTy}; |
| 4638 | return mlir::LLVM::LLVMStructType::getLiteral(context: type.getContext(), |
| 4639 | types: structFields); |
| 4640 | }); |
| 4641 | converter.addConversion(callback: [&](cir::FuncType type) -> std::optional<mlir::Type> { |
| 4642 | auto result = converter.convertType(t: type.getReturnType()); |
| 4643 | if (!result) |
| 4644 | return std::nullopt; |
| 4645 | llvm::SmallVector<mlir::Type> arguments; |
| 4646 | arguments.reserve(N: type.getNumInputs()); |
| 4647 | if (converter.convertTypes(types: type.getInputs(), results&: arguments).failed()) |
| 4648 | return std::nullopt; |
| 4649 | auto varArg = type.isVarArg(); |
| 4650 | return mlir::LLVM::LLVMFunctionType::get(result, arguments, isVarArg: varArg); |
| 4651 | }); |
| 4652 | converter.addConversion(callback: [&](cir::StructType type) -> mlir::Type { |
| 4653 | llvm::SmallVector<mlir::Type> llvmMembers; |
| 4654 | for (mlir::Type ty : type.getMembers()) { |
| 4655 | // A bit-field that shares an earlier field's access unit, and a |
| 4656 | // zero-width bit-field, own no bytes, so they get no field of their own. |
| 4657 | // StructType::getLLVMFieldIndex maps the two numberings. |
| 4658 | if (!cir::memberOwnsBytes(memberTy: ty)) |
| 4659 | continue; |
| 4660 | mlir::Type memberTy = convertTypeForMemory(converter, dataLayout, |
| 4661 | type: cir::memberStorageType(memberTy: ty)); |
| 4662 | // A null member means an unsupported type (e.g. a _BitInt with byte-array |
| 4663 | // storage); propagate the conversion failure instead of building an |
| 4664 | // invalid struct body. |
| 4665 | if (!memberTy) |
| 4666 | return {}; |
| 4667 | llvmMembers.push_back(Elt: memberTy); |
| 4668 | } |
| 4669 | |
| 4670 | mlir::LLVM::LLVMStructType llvmStruct; |
| 4671 | if (type.getName()) { |
| 4672 | llvmStruct = mlir::LLVM::LLVMStructType::getIdentified( |
| 4673 | context: type.getContext(), name: type.getPrefixedName()); |
| 4674 | if (llvmStruct.setBody(types: llvmMembers, isPacked: type.getPacked()).failed()) |
| 4675 | llvm_unreachable("Failed to set body of record" ); |
| 4676 | } else { |
| 4677 | llvmStruct = mlir::LLVM::LLVMStructType::getLiteral( |
| 4678 | context: type.getContext(), types: llvmMembers, isPacked: type.getPacked()); |
| 4679 | } |
| 4680 | return llvmStruct; |
| 4681 | }); |
| 4682 | // Unions are lowered as only the largest member. |
| 4683 | converter.addConversion(callback: [&](cir::UnionType type) -> mlir::Type { |
| 4684 | llvm::SmallVector<mlir::Type> llvmMembers; |
| 4685 | if (!type.getMembers().empty()) |
| 4686 | if (auto storage = type.getUnionStorageType(dataLayout)) { |
| 4687 | mlir::Type storageTy = |
| 4688 | convertTypeForMemory(converter, dataLayout, type: storage); |
| 4689 | if (!storageTy) |
| 4690 | return {}; |
| 4691 | llvmMembers.push_back(Elt: storageTy); |
| 4692 | } |
| 4693 | if (mlir::Type pad = type.getPadding()) { |
| 4694 | mlir::Type padTy = convertTypeForMemory(converter, dataLayout, type: pad); |
| 4695 | if (!padTy) |
| 4696 | return {}; |
| 4697 | llvmMembers.push_back(Elt: padTy); |
| 4698 | } |
| 4699 | |
| 4700 | mlir::LLVM::LLVMStructType llvmStruct; |
| 4701 | if (type.getName()) { |
| 4702 | llvmStruct = mlir::LLVM::LLVMStructType::getIdentified( |
| 4703 | context: type.getContext(), name: type.getPrefixedName()); |
| 4704 | if (llvmStruct.setBody(types: llvmMembers, isPacked: type.getPacked()).failed()) |
| 4705 | llvm_unreachable("Failed to set body of record" ); |
| 4706 | } else { |
| 4707 | llvmStruct = mlir::LLVM::LLVMStructType::getLiteral( |
| 4708 | context: type.getContext(), types: llvmMembers, isPacked: type.getPacked()); |
| 4709 | } |
| 4710 | return llvmStruct; |
| 4711 | }); |
| 4712 | converter.addConversion(callback: [&](cir::VoidType type) -> mlir::Type { |
| 4713 | return mlir::LLVM::LLVMVoidType::get(ctx: type.getContext()); |
| 4714 | }); |
| 4715 | } |
| 4716 | static void buildCtorDtorList( |
| 4717 | mlir::ModuleOp module, StringRef globalXtorName, StringRef llvmXtorName, |
| 4718 | llvm::function_ref<cir::GlobalCtorDtorEntry(mlir::Attribute)> createXtor) { |
| 4719 | llvm::SmallVector<cir::GlobalCtorDtorEntry> globalXtors; |
| 4720 | if (auto attr = |
| 4721 | module->getDiscardableAttrOfType<mlir::ArrayAttr>(name: globalXtorName)) |
| 4722 | for (mlir::Attribute element : attr) |
| 4723 | globalXtors.emplace_back(Args: createXtor(element)); |
| 4724 | |
| 4725 | if (globalXtors.empty()) |
| 4726 | return; |
| 4727 | |
| 4728 | mlir::OpBuilder builder(module.getContext()); |
| 4729 | builder.setInsertionPointToEnd(&module.getBodyRegion().back()); |
| 4730 | |
| 4731 | // Create a global array llvm.global_ctors with element type of |
| 4732 | // struct { i32, ptr, ptr } |
| 4733 | cir::CIRDataLayout cirDataLayout(module); |
| 4734 | auto ctorPFTy = mlir::LLVM::LLVMPointerType::get( |
| 4735 | context: builder.getContext(), addressSpace: cirDataLayout.getProgramAddressSpace()); |
| 4736 | llvm::SmallVector<mlir::Type> ctorStructFields; |
| 4737 | ctorStructFields.push_back(Elt: builder.getI32Type()); |
| 4738 | ctorStructFields.push_back(Elt: ctorPFTy); |
| 4739 | ctorStructFields.push_back(Elt: ctorPFTy); |
| 4740 | |
| 4741 | auto ctorStructTy = mlir::LLVM::LLVMStructType::getLiteral( |
| 4742 | context: builder.getContext(), types: ctorStructFields); |
| 4743 | auto ctorStructArrayTy = |
| 4744 | mlir::LLVM::LLVMArrayType::get(elementType: ctorStructTy, numElements: globalXtors.size()); |
| 4745 | |
| 4746 | mlir::Location loc = module.getLoc(); |
| 4747 | auto newGlobalOp = mlir::LLVM::GlobalOp::create( |
| 4748 | builder, location: loc, type: ctorStructArrayTy, /*constant=*/isConstant: false, |
| 4749 | linkage: mlir::LLVM::Linkage::Appending, name: llvmXtorName, value: mlir::Attribute()); |
| 4750 | |
| 4751 | builder.createBlock(parent: &newGlobalOp.getRegion()); |
| 4752 | builder.setInsertionPointToEnd(newGlobalOp.getInitializerBlock()); |
| 4753 | |
| 4754 | mlir::Value result = |
| 4755 | mlir::LLVM::UndefOp::create(builder, location: loc, res: ctorStructArrayTy); |
| 4756 | |
| 4757 | for (auto [index, xtorInfo] : llvm::enumerate(First&: globalXtors)) { |
| 4758 | mlir::Value structInit = |
| 4759 | mlir::LLVM::UndefOp::create(builder, location: loc, res: ctorStructTy); |
| 4760 | mlir::Value initPriority = mlir::LLVM::ConstantOp::create( |
| 4761 | builder, location: loc, type: ctorStructFields[0], value: xtorInfo.priority); |
| 4762 | mlir::Value initFuncAddr = mlir::LLVM::AddressOfOp::create( |
| 4763 | builder, location: loc, res: ctorStructFields[1], global_name: xtorInfo.name); |
| 4764 | mlir::Value initAssociate = |
| 4765 | xtorInfo.associated.empty() |
| 4766 | ? mlir::LLVM::ZeroOp::create(builder, location: loc, res: ctorStructFields[2]) |
| 4767 | .getResult() |
| 4768 | : mlir::LLVM::AddressOfOp::create(builder, location: loc, res: ctorStructFields[2], |
| 4769 | global_name: xtorInfo.associated) |
| 4770 | .getResult(); |
| 4771 | // Literal zero makes the InsertValueOp::create ambiguous. |
| 4772 | llvm::SmallVector<int64_t> zero{0}; |
| 4773 | structInit = mlir::LLVM::InsertValueOp::create(builder, location: loc, container: structInit, |
| 4774 | value: initPriority, position: zero); |
| 4775 | structInit = mlir::LLVM::InsertValueOp::create(builder, location: loc, container: structInit, |
| 4776 | value: initFuncAddr, position: 1); |
| 4777 | structInit = mlir::LLVM::InsertValueOp::create(builder, location: loc, container: structInit, |
| 4778 | value: initAssociate, position: 2); |
| 4779 | result = mlir::LLVM::InsertValueOp::create(builder, location: loc, container: result, value: structInit, |
| 4780 | position: index); |
| 4781 | } |
| 4782 | |
| 4783 | mlir::LLVM::ReturnOp::create(builder, location: loc, arg: result); |
| 4784 | } |
| 4785 | |
| 4786 | mlir::LogicalResult CIRToLLVMObjSizeOpLowering::matchAndRewrite( |
| 4787 | cir::ObjSizeOp op, OpAdaptor adaptor, |
| 4788 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4789 | mlir::Type llvmResTy = getTypeConverter()->convertType(t: op.getType()); |
| 4790 | mlir::Location loc = op->getLoc(); |
| 4791 | |
| 4792 | mlir::IntegerType i1Ty = rewriter.getI1Type(); |
| 4793 | |
| 4794 | auto i1Val = [&rewriter, &loc, &i1Ty](bool val) { |
| 4795 | return mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: i1Ty, value: val); |
| 4796 | }; |
| 4797 | |
| 4798 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm.objectsize" , resultTy: llvmResTy, |
| 4799 | operands: { |
| 4800 | adaptor.getPtr(), |
| 4801 | i1Val(op.getMin()), |
| 4802 | i1Val(op.getNullunknown()), |
| 4803 | i1Val(op.getDynamic()), |
| 4804 | }); |
| 4805 | |
| 4806 | return mlir::LogicalResult::success(); |
| 4807 | } |
| 4808 | |
| 4809 | //===----------------------------------------------------------------------===// |
| 4810 | // @llvm.global.annotations emission |
| 4811 | //===----------------------------------------------------------------------===// |
| 4812 | |
| 4813 | /// Annotation strings and arguments, shared by all annotations in a module. |
| 4814 | struct LLVMAnnotationGlobals { |
| 4815 | explicit LLVMAnnotationGlobals(mlir::ModuleOp module) { |
| 4816 | assert(!cir::MissingFeatures::annotationConstantAddrSpace()); |
| 4817 | auto asAttr = mlir::dyn_cast_if_present<mlir::IntegerAttr>( |
| 4818 | Val: mlir::DataLayout(module).getGlobalMemorySpace()); |
| 4819 | addrSpace = asAttr ? asAttr.getUInt() : 0; |
| 4820 | } |
| 4821 | |
| 4822 | unsigned addrSpace; |
| 4823 | llvm::StringMap<mlir::LLVM::GlobalOp> stringCache; |
| 4824 | llvm::StringMap<mlir::LLVM::GlobalOp> argStringCache; |
| 4825 | llvm::MapVector<mlir::ArrayAttr, mlir::LLVM::GlobalOp> argsCache; |
| 4826 | }; |
| 4827 | |
| 4828 | namespace { |
| 4829 | constexpr StringRef llvmMetadataSectionName = "llvm.metadata" ; |
| 4830 | |
| 4831 | /// Get-or-create a private constant string global in the llvm.metadata |
| 4832 | /// section, deduplicated by string content. |
| 4833 | mlir::LLVM::GlobalOp |
| 4834 | getOrCreateAnnotationStringGlobal(mlir::OpBuilder &builder, mlir::Location loc, |
| 4835 | mlir::ModuleOp module, llvm::StringRef str, |
| 4836 | llvm::StringMap<mlir::LLVM::GlobalOp> &cache, |
| 4837 | bool isArg, unsigned addrSpace) { |
| 4838 | auto it = cache.find(Key: str); |
| 4839 | if (it != cache.end()) |
| 4840 | return it->second; |
| 4841 | |
| 4842 | auto i8Ty = mlir::IntegerType::get(context: module.getContext(), width: 8); |
| 4843 | auto arrayTy = mlir::LLVM::LLVMArrayType::get(elementType: i8Ty, numElements: str.size() + 1); |
| 4844 | std::string name = ".str" ; |
| 4845 | if (!cache.empty()) |
| 4846 | name += "." + std::to_string(val: cache.size()); |
| 4847 | name += ".annotation" ; |
| 4848 | if (isArg) |
| 4849 | name += ".arg" ; |
| 4850 | |
| 4851 | mlir::LLVM::GlobalOp strGlobal = mlir::LLVM::GlobalOp::create( |
| 4852 | builder, location: loc, type: arrayTy, /*isConstant=*/true, linkage: mlir::LLVM::Linkage::Private, |
| 4853 | name, value: mlir::StringAttr::get(context: module.getContext(), bytes: std::string(str) + '\0'), |
| 4854 | /*alignment=*/isArg ? 1 : 0, addrSpace); |
| 4855 | if (!isArg) |
| 4856 | strGlobal.setSection(llvmMetadataSectionName); |
| 4857 | strGlobal.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Global); |
| 4858 | strGlobal.setDsoLocal(true); |
| 4859 | cache[str] = strGlobal; |
| 4860 | return strGlobal; |
| 4861 | } |
| 4862 | |
| 4863 | /// Get-or-create a private constant struct holding the annotation arguments, |
| 4864 | /// deduplicated by ArrayAttr identity. |
| 4865 | mlir::LLVM::GlobalOp getOrCreateAnnotationArgsVar( |
| 4866 | mlir::OpBuilder &builder, mlir::Location loc, mlir::ModuleOp module, |
| 4867 | mlir::ArrayAttr argsAttr, |
| 4868 | llvm::StringMap<mlir::LLVM::GlobalOp> &argStringCache, |
| 4869 | llvm::MapVector<mlir::ArrayAttr, mlir::LLVM::GlobalOp> &argsCache, |
| 4870 | unsigned addrSpace) { |
| 4871 | auto it = argsCache.find(Key: argsAttr); |
| 4872 | if (it != argsCache.end()) |
| 4873 | return it->second; |
| 4874 | |
| 4875 | auto ptrTy = |
| 4876 | mlir::LLVM::LLVMPointerType::get(context: builder.getContext(), addressSpace: addrSpace); |
| 4877 | |
| 4878 | llvm::SmallVector<mlir::Type> fieldTypes; |
| 4879 | for (mlir::Attribute arg : argsAttr) { |
| 4880 | if (mlir::isa<mlir::StringAttr>(Val: arg)) |
| 4881 | fieldTypes.push_back(Elt: ptrTy); |
| 4882 | else if (auto intAttr = mlir::dyn_cast<mlir::IntegerAttr>(Val&: arg)) |
| 4883 | fieldTypes.push_back(Elt: intAttr.getType()); |
| 4884 | else |
| 4885 | llvm_unreachable("Unsupported annotation arg type" ); |
| 4886 | } |
| 4887 | |
| 4888 | auto structTy = |
| 4889 | mlir::LLVM::LLVMStructType::getLiteral(context: builder.getContext(), types: fieldTypes); |
| 4890 | std::string name = ".args" ; |
| 4891 | if (!argsCache.empty()) |
| 4892 | name += "." + std::to_string(val: argsCache.size()); |
| 4893 | name += ".annotation" ; |
| 4894 | |
| 4895 | mlir::LLVM::GlobalOp argsGlobal = mlir::LLVM::GlobalOp::create( |
| 4896 | builder, location: loc, type: structTy, /*isConstant=*/true, linkage: mlir::LLVM::Linkage::Private, |
| 4897 | name, value: mlir::Attribute(), /*alignment=*/0, addrSpace); |
| 4898 | argsGlobal.setSection(llvmMetadataSectionName); |
| 4899 | argsGlobal.setUnnamedAddr(mlir::LLVM::UnnamedAddr::Global); |
| 4900 | argsGlobal.setDsoLocal(true); |
| 4901 | |
| 4902 | // Build the initializer block. |
| 4903 | argsGlobal.getRegion().push_back(block: new mlir::Block()); |
| 4904 | mlir::OpBuilder initBuilder(module.getContext()); |
| 4905 | initBuilder.setInsertionPointToEnd(argsGlobal.getInitializerBlock()); |
| 4906 | |
| 4907 | mlir::Value structInit = |
| 4908 | mlir::LLVM::UndefOp::create(builder&: initBuilder, location: loc, res: structTy); |
| 4909 | for (auto [idx, arg] : llvm::enumerate(First&: argsAttr)) { |
| 4910 | if (auto strArg = mlir::dyn_cast<mlir::StringAttr>(Val: arg)) { |
| 4911 | mlir::LLVM::GlobalOp strGlobal = getOrCreateAnnotationStringGlobal( |
| 4912 | builder, loc, module, str: strArg.getValue(), cache&: argStringCache, |
| 4913 | /*isArg=*/true, addrSpace); |
| 4914 | mlir::LLVM::AddressOfOp strAddr = mlir::LLVM::AddressOfOp::create( |
| 4915 | builder&: initBuilder, location: loc, res: ptrTy, global_name: strGlobal.getSymName()); |
| 4916 | structInit = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: loc, |
| 4917 | container: structInit, value: strAddr, position: idx); |
| 4918 | } else if (auto intArg = mlir::dyn_cast<mlir::IntegerAttr>(Val: arg)) { |
| 4919 | mlir::LLVM::ConstantOp intConst = mlir::LLVM::ConstantOp::create( |
| 4920 | builder&: initBuilder, location: loc, type: intArg.getType(), value: intArg.getValue()); |
| 4921 | structInit = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: loc, |
| 4922 | container: structInit, value: intConst, position: idx); |
| 4923 | } else { |
| 4924 | llvm_unreachable("Unsupported annotation arg type" ); |
| 4925 | } |
| 4926 | } |
| 4927 | mlir::LLVM::ReturnOp::create(builder&: initBuilder, location: loc, arg: structInit); |
| 4928 | |
| 4929 | argsCache[argsAttr] = argsGlobal; |
| 4930 | return argsGlobal; |
| 4931 | } |
| 4932 | |
| 4933 | /// Resolve a possibly-fused MLIR Location to a FileLineColLoc, returning |
| 4934 | /// {filename, line}. Returns {empty, 0} if no usable file location is found. |
| 4935 | std::pair<llvm::StringRef, unsigned> (mlir::Location loc) { |
| 4936 | mlir::Location resolved = loc; |
| 4937 | if (auto fused = mlir::dyn_cast<mlir::FusedLoc>(Val&: resolved)) { |
| 4938 | if (!fused.getLocations().empty()) |
| 4939 | resolved = fused.getLocations()[0]; |
| 4940 | } |
| 4941 | if (auto fl = mlir::dyn_cast<mlir::FileLineColLoc>(Val&: resolved)) |
| 4942 | return {fl.getFilename().getValue(), fl.getLine()}; |
| 4943 | return {"" , 0}; |
| 4944 | } |
| 4945 | |
| 4946 | /// The name, file, line, and args operands of the llvm.*annotation intrinsics. |
| 4947 | struct AnnotationOperands { |
| 4948 | mlir::Value name, file, line, args; |
| 4949 | }; |
| 4950 | AnnotationOperands getAnnotationOperands(mlir::OpBuilder &builder, |
| 4951 | mlir::Location loc, |
| 4952 | cir::AnnotationAttr annotation, |
| 4953 | LLVMAnnotationGlobals &globals) { |
| 4954 | auto module = builder.getInsertionBlock() |
| 4955 | ->getParentOp() |
| 4956 | ->getParentOfType<mlir::ModuleOp>(); |
| 4957 | auto [filename, line] = extractFileLine(loc); |
| 4958 | mlir::ArrayAttr args = annotation.getArgs(); |
| 4959 | mlir::LLVM::GlobalOp nameGlobal, fileGlobal, argsGlobal; |
| 4960 | { |
| 4961 | mlir::OpBuilder::InsertionGuard guard(builder); |
| 4962 | builder.setInsertionPointToEnd(module.getBody()); |
| 4963 | nameGlobal = getOrCreateAnnotationStringGlobal( |
| 4964 | builder, loc, module, str: annotation.getName().getValue(), |
| 4965 | cache&: globals.stringCache, /*isArg=*/false, addrSpace: globals.addrSpace); |
| 4966 | fileGlobal = getOrCreateAnnotationStringGlobal( |
| 4967 | builder, loc, module, str: filename, cache&: globals.stringCache, /*isArg=*/false, |
| 4968 | addrSpace: globals.addrSpace); |
| 4969 | if (args && !args.empty()) |
| 4970 | argsGlobal = getOrCreateAnnotationArgsVar( |
| 4971 | builder, loc, module, argsAttr: args, argStringCache&: globals.argStringCache, argsCache&: globals.argsCache, |
| 4972 | addrSpace: globals.addrSpace); |
| 4973 | } |
| 4974 | |
| 4975 | auto ptrTy = |
| 4976 | mlir::LLVM::LLVMPointerType::get(context: builder.getContext(), addressSpace: globals.addrSpace); |
| 4977 | AnnotationOperands ops; |
| 4978 | ops.name = mlir::LLVM::AddressOfOp::create(builder, location: loc, res: ptrTy, |
| 4979 | global_name: nameGlobal.getSymName()); |
| 4980 | ops.file = mlir::LLVM::AddressOfOp::create(builder, location: loc, res: ptrTy, |
| 4981 | global_name: fileGlobal.getSymName()); |
| 4982 | ops.line = |
| 4983 | mlir::LLVM::ConstantOp::create(builder, location: loc, type: builder.getI32Type(), value: line); |
| 4984 | if (argsGlobal) |
| 4985 | ops.args = mlir::LLVM::AddressOfOp::create(builder, location: loc, res: ptrTy, |
| 4986 | global_name: argsGlobal.getSymName()); |
| 4987 | else |
| 4988 | ops.args = mlir::LLVM::ZeroOp::create(builder, location: loc, res: ptrTy); |
| 4989 | return ops; |
| 4990 | } |
| 4991 | } // namespace |
| 4992 | |
| 4993 | mlir::LogicalResult CIRToLLVMVarAnnotationOpLowering::matchAndRewrite( |
| 4994 | cir::VarAnnotationOp op, OpAdaptor adaptor, |
| 4995 | mlir::ConversionPatternRewriter &rewriter) const { |
| 4996 | AnnotationOperands ops = getAnnotationOperands( |
| 4997 | builder&: rewriter, loc: op.getLoc(), annotation: op.getAnnotation(), globals&: annotationGlobals); |
| 4998 | rewriter.replaceOpWithNewOp<mlir::LLVM::VarAnnotation>( |
| 4999 | op, args: adaptor.getAddr(), args&: ops.name, args&: ops.file, args&: ops.line, args&: ops.args); |
| 5000 | return mlir::success(); |
| 5001 | } |
| 5002 | |
| 5003 | mlir::LogicalResult CIRToLLVMPtrAnnotationOpLowering::matchAndRewrite( |
| 5004 | cir::PtrAnnotationOp op, OpAdaptor adaptor, |
| 5005 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5006 | AnnotationOperands ops = getAnnotationOperands( |
| 5007 | builder&: rewriter, loc: op.getLoc(), annotation: op.getAnnotation(), globals&: annotationGlobals); |
| 5008 | rewriter.replaceOpWithNewOp<mlir::LLVM::PtrAnnotation>( |
| 5009 | op, args: adaptor.getAddr().getType(), args: adaptor.getAddr(), args&: ops.name, args&: ops.file, |
| 5010 | args&: ops.line, args&: ops.args); |
| 5011 | return mlir::success(); |
| 5012 | } |
| 5013 | |
| 5014 | mlir::LogicalResult CIRToLLVMAnnotationOpLowering::matchAndRewrite( |
| 5015 | cir::AnnotationOp op, OpAdaptor adaptor, |
| 5016 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5017 | AnnotationOperands ops = getAnnotationOperands( |
| 5018 | builder&: rewriter, loc: op.getLoc(), annotation: op.getAnnotation(), globals&: annotationGlobals); |
| 5019 | rewriter.replaceOpWithNewOp<mlir::LLVM::Annotation>( |
| 5020 | op, args: adaptor.getValue().getType(), args: adaptor.getValue(), args&: ops.name, args&: ops.file, |
| 5021 | args&: ops.line); |
| 5022 | return mlir::success(); |
| 5023 | } |
| 5024 | |
| 5025 | void ConvertCIRToLLVMPass::collectGlobalAnnotations(mlir::ModuleOp module) { |
| 5026 | auto handleArray = [&](mlir::StringAttr symName, mlir::ArrayAttr arr, |
| 5027 | mlir::Location loc) { |
| 5028 | if (!arr) |
| 5029 | return; |
| 5030 | for (mlir::Attribute a : arr) |
| 5031 | if (auto annot = mlir::dyn_cast<cir::AnnotationAttr>(Val&: a)) |
| 5032 | collectedAnnotations.emplace_back(Args&: symName, Args&: annot, Args&: loc); |
| 5033 | }; |
| 5034 | |
| 5035 | // Walk in IR order: GlobalOps first (they appear before functions in the |
| 5036 | // module body), then FuncOps. This matches OGCG's emission order. |
| 5037 | module.walk(callback: [&](cir::GlobalOp op) { |
| 5038 | handleArray(op.getSymNameAttr(), op.getAnnotationsAttr(), op.getLoc()); |
| 5039 | }); |
| 5040 | module.walk(callback: [&](cir::FuncOp op) { |
| 5041 | handleArray(op.getSymNameAttr(), op.getAnnotationsAttr(), op.getLoc()); |
| 5042 | }); |
| 5043 | } |
| 5044 | |
| 5045 | void ConvertCIRToLLVMPass::buildGlobalAnnotationsVar( |
| 5046 | mlir::ModuleOp module, LLVMAnnotationGlobals &annotationGlobals) { |
| 5047 | if (collectedAnnotations.empty()) |
| 5048 | return; |
| 5049 | |
| 5050 | mlir::MLIRContext *ctx = module.getContext(); |
| 5051 | mlir::OpBuilder builder(ctx); |
| 5052 | builder.setInsertionPointToEnd(&module.getBodyRegion().back()); |
| 5053 | |
| 5054 | unsigned addrSpace = annotationGlobals.addrSpace; |
| 5055 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: ctx, addressSpace: addrSpace); |
| 5056 | auto i32Ty = builder.getI32Type(); |
| 5057 | |
| 5058 | // Each entry: { ptr, ptr, ptr, i32, ptr }. |
| 5059 | auto entryTy = mlir::LLVM::LLVMStructType::getLiteral( |
| 5060 | context: ctx, types: {ptrTy, ptrTy, ptrTy, i32Ty, ptrTy}); |
| 5061 | auto arrayTy = |
| 5062 | mlir::LLVM::LLVMArrayType::get(elementType: entryTy, numElements: collectedAnnotations.size()); |
| 5063 | |
| 5064 | mlir::Location moduleLoc = module.getLoc(); |
| 5065 | auto annotationsGlobal = mlir::LLVM::GlobalOp::create( |
| 5066 | builder, location: moduleLoc, type: arrayTy, /*isConstant=*/false, |
| 5067 | linkage: mlir::LLVM::Linkage::Appending, name: "llvm.global.annotations" , |
| 5068 | value: mlir::Attribute(), /*alignment=*/0, addrSpace); |
| 5069 | annotationsGlobal.setSection(llvmMetadataSectionName); |
| 5070 | |
| 5071 | // Strings/args constants must come *before* @llvm.global.annotations to |
| 5072 | // match OGCG output order. Insert them just before the annotations global. |
| 5073 | mlir::OpBuilder constsBuilder(ctx); |
| 5074 | constsBuilder.setInsertionPoint(annotationsGlobal); |
| 5075 | |
| 5076 | llvm::StringMap<mlir::LLVM::GlobalOp> &stringCache = |
| 5077 | annotationGlobals.stringCache; |
| 5078 | llvm::StringMap<mlir::LLVM::GlobalOp> &argStringCache = |
| 5079 | annotationGlobals.argStringCache; |
| 5080 | llvm::MapVector<mlir::ArrayAttr, mlir::LLVM::GlobalOp> &argsCache = |
| 5081 | annotationGlobals.argsCache; |
| 5082 | |
| 5083 | // Build the initializer block of @llvm.global.annotations. |
| 5084 | annotationsGlobal.getRegion().push_back(block: new mlir::Block()); |
| 5085 | mlir::OpBuilder initBuilder(ctx); |
| 5086 | initBuilder.setInsertionPointToEnd(annotationsGlobal.getInitializerBlock()); |
| 5087 | |
| 5088 | mlir::Value arrayVal = |
| 5089 | mlir::LLVM::UndefOp::create(builder&: initBuilder, location: moduleLoc, res: arrayTy); |
| 5090 | |
| 5091 | for (auto [idx, entry] : llvm::enumerate(First&: collectedAnnotations)) { |
| 5092 | mlir::Value entryVal = |
| 5093 | mlir::LLVM::UndefOp::create(builder&: initBuilder, location: moduleLoc, res: entryTy); |
| 5094 | |
| 5095 | // Field 0: ptr to the annotated symbol. (Literal zero is ambiguous on |
| 5096 | // InsertValueOp::create, wrap in a SmallVector.) |
| 5097 | llvm::SmallVector<int64_t> zero{0}; |
| 5098 | unsigned symAddrSpace = 0; |
| 5099 | if (auto global = mlir::dyn_cast_if_present<mlir::LLVM::GlobalOp>( |
| 5100 | Val: mlir::SymbolTable::lookupSymbolIn(op: module, symbol: entry.symName))) |
| 5101 | symAddrSpace = global.getAddrSpace(); |
| 5102 | mlir::Value symAddr = mlir::LLVM::AddressOfOp::create( |
| 5103 | builder&: initBuilder, location: moduleLoc, |
| 5104 | res: mlir::LLVM::LLVMPointerType::get(context: ctx, addressSpace: symAddrSpace), |
| 5105 | global_name: entry.symName.getValue()); |
| 5106 | if (symAddrSpace != addrSpace) |
| 5107 | symAddr = mlir::LLVM::AddrSpaceCastOp::create(builder&: initBuilder, location: moduleLoc, |
| 5108 | res: ptrTy, arg: symAddr); |
| 5109 | entryVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5110 | container: entryVal, value: symAddr, position: zero); |
| 5111 | |
| 5112 | // Field 1: ptr to the annotation name string. |
| 5113 | mlir::LLVM::GlobalOp nameGlobal = getOrCreateAnnotationStringGlobal( |
| 5114 | builder&: constsBuilder, loc: moduleLoc, module, str: entry.annotation.getName().getValue(), |
| 5115 | cache&: stringCache, /*isArg=*/false, addrSpace); |
| 5116 | mlir::LLVM::AddressOfOp nameAddr = mlir::LLVM::AddressOfOp::create( |
| 5117 | builder&: initBuilder, location: moduleLoc, res: ptrTy, global_name: nameGlobal.getSymName()); |
| 5118 | entryVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5119 | container: entryVal, value: nameAddr, position: 1); |
| 5120 | |
| 5121 | // Fields 2 and 3: ptr to filename string and line number. |
| 5122 | auto [filename, line] = extractFileLine(loc: entry.loc); |
| 5123 | mlir::LLVM::GlobalOp fileGlobal = getOrCreateAnnotationStringGlobal( |
| 5124 | builder&: constsBuilder, loc: moduleLoc, module, str: filename, cache&: stringCache, |
| 5125 | /*isArg=*/false, addrSpace); |
| 5126 | mlir::LLVM::AddressOfOp fileAddr = mlir::LLVM::AddressOfOp::create( |
| 5127 | builder&: initBuilder, location: moduleLoc, res: ptrTy, global_name: fileGlobal.getSymName()); |
| 5128 | entryVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5129 | container: entryVal, value: fileAddr, position: 2); |
| 5130 | mlir::LLVM::ConstantOp lineConst = |
| 5131 | mlir::LLVM::ConstantOp::create(builder&: initBuilder, location: moduleLoc, type: i32Ty, value: line); |
| 5132 | entryVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5133 | container: entryVal, value: lineConst, position: 3); |
| 5134 | |
| 5135 | // Field 4: ptr to args, or null if none. |
| 5136 | mlir::ArrayAttr args = entry.annotation.getArgs(); |
| 5137 | mlir::Value argsField; |
| 5138 | if (!args || args.empty()) { |
| 5139 | argsField = mlir::LLVM::ZeroOp::create(builder&: initBuilder, location: moduleLoc, res: ptrTy); |
| 5140 | } else { |
| 5141 | mlir::LLVM::GlobalOp argsGlobal = |
| 5142 | getOrCreateAnnotationArgsVar(builder&: constsBuilder, loc: moduleLoc, module, argsAttr: args, |
| 5143 | argStringCache, argsCache, addrSpace); |
| 5144 | argsField = mlir::LLVM::AddressOfOp::create(builder&: initBuilder, location: moduleLoc, res: ptrTy, |
| 5145 | global_name: argsGlobal.getSymName()); |
| 5146 | } |
| 5147 | entryVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5148 | container: entryVal, value: argsField, position: 4); |
| 5149 | |
| 5150 | arrayVal = mlir::LLVM::InsertValueOp::create(builder&: initBuilder, location: moduleLoc, |
| 5151 | container: arrayVal, value: entryVal, position: idx); |
| 5152 | } |
| 5153 | |
| 5154 | mlir::LLVM::ReturnOp::create(builder&: initBuilder, location: moduleLoc, arg: arrayVal); |
| 5155 | } |
| 5156 | |
| 5157 | void ConvertCIRToLLVMPass::resolveBlockAddressOp( |
| 5158 | LLVMBlockAddressInfo &blockInfoAddr) { |
| 5159 | |
| 5160 | mlir::ModuleOp module = getOperation(); |
| 5161 | mlir::OpBuilder opBuilder(module.getContext()); |
| 5162 | for (auto &[blockAddOp, blockInfo] : |
| 5163 | blockInfoAddr.getUnresolvedBlockAddress()) { |
| 5164 | mlir::LLVM::BlockTagOp resolvedLabel = |
| 5165 | blockInfoAddr.lookupBlockTag(info: blockInfo); |
| 5166 | assert(resolvedLabel && "expected BlockTagOp to already be emitted" ); |
| 5167 | mlir::FlatSymbolRefAttr fnSym = blockInfo.getFunc(); |
| 5168 | auto blkAddTag = mlir::LLVM::BlockAddressAttr::get( |
| 5169 | context: opBuilder.getContext(), function: fnSym, tag: resolvedLabel.getTagAttr()); |
| 5170 | blockAddOp.setBlockAddrAttr(blkAddTag); |
| 5171 | } |
| 5172 | blockInfoAddr.clearUnresolvedMap(); |
| 5173 | } |
| 5174 | |
| 5175 | void ConvertCIRToLLVMPass::processCIRAttrs(mlir::ModuleOp module) { |
| 5176 | // Lower the module attributes to LLVM equivalents. |
| 5177 | if (mlir::Attribute tripleAttr = |
| 5178 | module->getAttr(name: cir::CIRDialect::getTripleAttrName())) |
| 5179 | module->setAttr(name: mlir::LLVM::LLVMDialect::getTargetTripleAttrName(), |
| 5180 | value: tripleAttr); |
| 5181 | |
| 5182 | if (mlir::Attribute asmAttr = |
| 5183 | module->getAttr(name: cir::CIRDialect::getModuleLevelAsmAttrName())) |
| 5184 | module->setAttr(name: mlir::LLVM::LLVMDialect::getModuleLevelAsmAttrName(), |
| 5185 | value: asmAttr); |
| 5186 | |
| 5187 | if (auto framePointerAttr = module->getAttrOfType<cir::FramePointerKindAttr>( |
| 5188 | name: cir::CIRDialect::getModuleFramePointerAttrName())) { |
| 5189 | mlir::OpBuilder builder(module.getContext()); |
| 5190 | builder.setInsertionPointToEnd(module.getBody()); |
| 5191 | auto flag = mlir::LLVM::ModuleFlagAttr::get( |
| 5192 | context: module.getContext(), behavior: mlir::LLVM::ModFlagBehavior::Max, |
| 5193 | key: builder.getStringAttr(bytes: "frame-pointer" ), |
| 5194 | value: builder.getI32IntegerAttr(value: static_cast<int32_t>( |
| 5195 | convertFramePointerKindToLLVMIR(kind: framePointerAttr.getValue())))); |
| 5196 | mlir::LLVM::ModuleFlagsOp::create(builder, location: module.getLoc(), |
| 5197 | flags: builder.getArrayAttr(value: {flag})); |
| 5198 | } |
| 5199 | |
| 5200 | lowerOpenCLModuleMetadataAttrs(module); |
| 5201 | } |
| 5202 | |
| 5203 | void ConvertCIRToLLVMPass::runOnOperation() { |
| 5204 | llvm::TimeTraceScope scope("Convert CIR to LLVM Pass" ); |
| 5205 | |
| 5206 | mlir::ModuleOp module = getOperation(); |
| 5207 | // CIR type of each symbol before lowering, for lookups that find the |
| 5208 | // already-lowered LLVM symbol after replacement. This must cover every |
| 5209 | // symbol kind that can be replaced: a kind missing from this loop |
| 5210 | // quietly falls back to the adjusted LLVM type in readers (here and in |
| 5211 | // LoweringHelpers). |
| 5212 | // Aliases are covered: they are GlobalOp and FuncOp ops with an |
| 5213 | // aliasee, not separate symbol kinds. |
| 5214 | CIRPreConversionTypeMap preConversionTypes; |
| 5215 | for (mlir::Operation &op : module.getBody()->getOperations()) { |
| 5216 | if (auto globalOp = mlir::dyn_cast<cir::GlobalOp>(Val&: op)) |
| 5217 | preConversionTypes[globalOp.getSymName()] = globalOp.getSymType(); |
| 5218 | else if (auto funcOp = mlir::dyn_cast<cir::FuncOp>(Val&: op)) |
| 5219 | preConversionTypes[funcOp.getSymName()] = funcOp.getFunctionType(); |
| 5220 | } |
| 5221 | mlir::DataLayout dl(module); |
| 5222 | mlir::LLVMTypeConverter converter(&getContext()); |
| 5223 | prepareTypeConverter(converter, dataLayout&: dl); |
| 5224 | |
| 5225 | /// Tracks the state required to lower CIR `LabelOp` and `BlockAddressOp`. |
| 5226 | /// Maps labels to their corresponding `BlockTagOp` and keeps bookkeeping |
| 5227 | /// of unresolved `BlockAddressOp`s until they are matched with the |
| 5228 | /// corresponding `BlockTagOp` in `resolveBlockAddressOp`. |
| 5229 | LLVMBlockAddressInfo blockInfoAddr; |
| 5230 | /// Cached symbol table collection used by call lowering patterns to avoid |
| 5231 | /// repeated O(M) module-wide symbol scans for every call site. |
| 5232 | mlir::SymbolTableCollection symbolTables; |
| 5233 | LLVMAnnotationGlobals annotationGlobals(module); |
| 5234 | mlir::RewritePatternSet patterns(&getContext()); |
| 5235 | patterns.add<CIRToLLVMVarAnnotationOpLowering, |
| 5236 | CIRToLLVMPtrAnnotationOpLowering, CIRToLLVMAnnotationOpLowering>( |
| 5237 | arg&: converter, args: patterns.getContext(), args&: dl, args&: symbolTables, args&: annotationGlobals); |
| 5238 | patterns.add<CIRToLLVMBlockAddressOpLowering, CIRToLLVMLabelOpLowering>( |
| 5239 | arg&: converter, args: patterns.getContext(), args&: dl, args&: symbolTables, args&: blockInfoAddr); |
| 5240 | patterns.add<CIRToLLVMGlobalOpLowering, CIRToLLVMConstantOpLowering>( |
| 5241 | arg&: converter, args: patterns.getContext(), args&: dl, args&: symbolTables, args&: blockInfoAddr, |
| 5242 | args&: preConversionTypes); |
| 5243 | |
| 5244 | patterns.add< |
| 5245 | #define GET_LLVM_LOWERING_PATTERNS_LIST |
| 5246 | #include "clang/CIR/Dialect/IR/CIRLowering.inc" |
| 5247 | #undef GET_LLVM_LOWERING_PATTERNS_LIST |
| 5248 | >(arg&: converter, args: patterns.getContext(), args&: dl, args&: symbolTables); |
| 5249 | // Patterns reading replaced symbols take the pre-conversion type map. |
| 5250 | patterns |
| 5251 | .add<CIRToLLVMVTableAddrPointOpLowering, CIRToLLVMVTTAddrPointOpLowering>( |
| 5252 | arg&: converter, args: patterns.getContext(), args&: dl, args&: symbolTables, |
| 5253 | args&: preConversionTypes); |
| 5254 | |
| 5255 | processCIRAttrs(module); |
| 5256 | |
| 5257 | // Collect annotation info from cir.func / cir.global before conversion; |
| 5258 | // the annotations attribute is filtered out during FuncOp/GlobalOp lowering. |
| 5259 | collectGlobalAnnotations(module); |
| 5260 | |
| 5261 | // Resolve constant printf strings while their producers are still CIR. |
| 5262 | annotateAMDGPUPrintfStrings(module, symbolTables); |
| 5263 | |
| 5264 | mlir::ConversionTarget target(getContext()); |
| 5265 | target.addLegalOp<mlir::ModuleOp>(); |
| 5266 | target.addLegalDialect<mlir::LLVM::LLVMDialect>(); |
| 5267 | mlir::configureOpenMPToLLVMConversionLegality(target, typeConverter: converter); |
| 5268 | target.addLegalDialect<mlir::omp::OpenMPDialect>(); |
| 5269 | mlir::populateOpenMPToLLVMConversionPatterns(converter, patterns); |
| 5270 | target.addIllegalDialect<mlir::BuiltinDialect, cir::CIRDialect, |
| 5271 | mlir::func::FuncDialect>(); |
| 5272 | |
| 5273 | llvm::SmallVector<mlir::Operation *> ops; |
| 5274 | ops.push_back(Elt: module); |
| 5275 | cir::collectUnreachable(parent: module, ops); |
| 5276 | |
| 5277 | // One-shot conversion (no pattern rollback): rewrites materialize |
| 5278 | // immediately, skipping the IRRewrite/value-mapping bookkeeping. The |
| 5279 | // CIR patterns here are compatible: nothing dereferences the root op |
| 5280 | // after replacement, and the cached symbol table is maintained |
| 5281 | // explicitly at every module-symbol replacement. |
| 5282 | mlir::ConversionConfig config; |
| 5283 | config.allowPatternRollback = false; |
| 5284 | if (failed( |
| 5285 | Result: applyPartialConversion(ops, target, patterns: std::move(patterns), config))) { |
| 5286 | signalPassFailure(); |
| 5287 | return; |
| 5288 | } |
| 5289 | |
| 5290 | // Drop CIR-native data-layout entries. |
| 5291 | if (auto dlSpec = mlir::dyn_cast_or_null<mlir::DataLayoutSpecAttr>( |
| 5292 | Val: module->getAttr(name: mlir::DLTIDialect::kDataLayoutAttrName))) { |
| 5293 | llvm::SmallVector<mlir::DataLayoutEntryInterface> kept; |
| 5294 | for (mlir::DataLayoutEntryInterface entry : dlSpec.getEntries()) { |
| 5295 | if (entry.isTypeEntry()) { |
| 5296 | mlir::Type key = mlir::cast<mlir::Type>(Val: entry.getKey()); |
| 5297 | if (mlir::isa<cir::PointerType, cir::IntType, cir::VectorType>(Val: key)) |
| 5298 | continue; |
| 5299 | } |
| 5300 | kept.push_back(Elt: entry); |
| 5301 | } |
| 5302 | module->setAttr(name: mlir::DLTIDialect::kDataLayoutAttrName, |
| 5303 | value: mlir::DataLayoutSpecAttr::get(context: module.getContext(), entries: kept)); |
| 5304 | } |
| 5305 | |
| 5306 | // Emit the llvm.global_ctors array. |
| 5307 | buildCtorDtorList( |
| 5308 | module, globalXtorName: cir::CIRDialect::getGlobalCtorsAttrName(), llvmXtorName: "llvm.global_ctors" , |
| 5309 | createXtor: [](mlir::Attribute attr) { |
| 5310 | auto ctorAttr = mlir::cast<cir::GlobalCtorAttr>(Val&: attr); |
| 5311 | StringRef associated; |
| 5312 | if (mlir::FlatSymbolRefAttr assoc = ctorAttr.getAssociated()) |
| 5313 | associated = assoc.getValue(); |
| 5314 | return cir::GlobalCtorDtorEntry(ctorAttr.getName(), |
| 5315 | ctorAttr.getPriority(), associated); |
| 5316 | }); |
| 5317 | // Emit the llvm.global_dtors array. |
| 5318 | buildCtorDtorList( |
| 5319 | module, globalXtorName: cir::CIRDialect::getGlobalDtorsAttrName(), llvmXtorName: "llvm.global_dtors" , |
| 5320 | createXtor: [](mlir::Attribute attr) { |
| 5321 | auto dtorAttr = mlir::cast<cir::GlobalDtorAttr>(Val&: attr); |
| 5322 | StringRef associated; |
| 5323 | if (mlir::FlatSymbolRefAttr assoc = dtorAttr.getAssociated()) |
| 5324 | associated = assoc.getValue(); |
| 5325 | return cir::GlobalCtorDtorEntry(dtorAttr.getName(), |
| 5326 | dtorAttr.getPriority(), associated); |
| 5327 | }); |
| 5328 | // Emit @llvm.global.annotations from the previously-collected entries. |
| 5329 | buildGlobalAnnotationsVar(module, annotationGlobals); |
| 5330 | |
| 5331 | resolveBlockAddressOp(blockInfoAddr); |
| 5332 | } |
| 5333 | |
| 5334 | mlir::LogicalResult CIRToLLVMBrOpLowering::matchAndRewrite( |
| 5335 | cir::BrOp op, OpAdaptor adaptor, |
| 5336 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5337 | rewriter.replaceOpWithNewOp<mlir::LLVM::BrOp>(op, args: adaptor.getOperands(), |
| 5338 | args: op.getDest()); |
| 5339 | return mlir::LogicalResult::success(); |
| 5340 | } |
| 5341 | |
| 5342 | mlir::LogicalResult CIRToLLVMGetMemberOpLowering::matchAndRewrite( |
| 5343 | cir::GetMemberOp op, OpAdaptor adaptor, |
| 5344 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5345 | mlir::Type llResTy = getTypeConverter()->convertType(t: op.getType()); |
| 5346 | if (!llResTy) |
| 5347 | return emitTypeConversionError(op, type: op.getType()); |
| 5348 | mlir::Type pointee = op.getAddrTy().getPointee(); |
| 5349 | |
| 5350 | if (mlir::isa<cir::UnionType>(Val: pointee)) { |
| 5351 | // Union members share the address space, so we just need a bitcast to |
| 5352 | // conform to type-checking. |
| 5353 | rewriter.replaceOpWithNewOp<mlir::LLVM::BitcastOp>(op, args&: llResTy, |
| 5354 | args: adaptor.getAddr()); |
| 5355 | return mlir::success(); |
| 5356 | } |
| 5357 | |
| 5358 | auto structTy = mlir::cast<cir::StructType>(Val&: pointee); |
| 5359 | // Since the base address is a pointer to an aggregate, the first offset |
| 5360 | // is always zero. The second offset tells us which member it will access. |
| 5361 | // Members that own no bytes are left out of the LLVM body, so the CIR member |
| 5362 | // index has to be mapped rather than used directly. |
| 5363 | llvm::SmallVector<mlir::LLVM::GEPArg, 2> offset{ |
| 5364 | 0, static_cast<int32_t>(structTy.getLLVMFieldIndex(idx: op.getIndex()))}; |
| 5365 | const mlir::Type elementTy = getTypeConverter()->convertType(t: structTy); |
| 5366 | if (!elementTy) |
| 5367 | return emitTypeConversionError(op, type: structTy); |
| 5368 | // Struct member accesses are always inbounds and nuw: the base pointer |
| 5369 | // is valid and the member offset is a positive, constant offset within |
| 5370 | // the struct layout, so it cannot wrap. This matches LLVM's |
| 5371 | // IRBuilder::CreateStructGEP. |
| 5372 | mlir::LLVM::GEPNoWrapFlags flags = |
| 5373 | mlir::LLVM::GEPNoWrapFlags::inbounds | mlir::LLVM::GEPNoWrapFlags::nuw; |
| 5374 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 5375 | op, args&: llResTy, args: elementTy, args: adaptor.getAddr(), args&: offset, args&: flags); |
| 5376 | return mlir::success(); |
| 5377 | } |
| 5378 | |
| 5379 | mlir::LogicalResult CIRToLLVMExtractMemberOpLowering::matchAndRewrite( |
| 5380 | cir::ExtractMemberOp op, OpAdaptor adaptor, |
| 5381 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5382 | if (mlir::isa<cir::UnionType>(Val: op.getRecord().getType())) { |
| 5383 | op.emitError(message: "cir.extract_member cannot extract member from a union" ); |
| 5384 | return mlir::failure(); |
| 5385 | } |
| 5386 | |
| 5387 | auto structTy = mlir::cast<cir::StructType>(Val: op.getRecord().getType()); |
| 5388 | std::int64_t indices[1] = { |
| 5389 | static_cast<std::int64_t>(structTy.getLLVMFieldIndex(idx: op.getIndex()))}; |
| 5390 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractValueOp>( |
| 5391 | op, args: adaptor.getRecord(), args&: indices); |
| 5392 | return mlir::success(); |
| 5393 | } |
| 5394 | |
| 5395 | mlir::LogicalResult CIRToLLVMInsertMemberOpLowering::matchAndRewrite( |
| 5396 | cir::InsertMemberOp op, OpAdaptor adaptor, |
| 5397 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5398 | if (mlir::isa<cir::UnionType>(Val: op.getRecord().getType())) { |
| 5399 | op.emitError(message: "cir.update_member cannot update member of a union" ); |
| 5400 | return mlir::failure(); |
| 5401 | } |
| 5402 | |
| 5403 | auto structTy = mlir::cast<cir::StructType>(Val: op.getRecord().getType()); |
| 5404 | std::int64_t indices[1] = { |
| 5405 | static_cast<std::int64_t>(structTy.getLLVMFieldIndex(idx: op.getIndex()))}; |
| 5406 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
| 5407 | op, args: adaptor.getRecord(), args: adaptor.getValue(), args&: indices); |
| 5408 | return mlir::success(); |
| 5409 | } |
| 5410 | |
| 5411 | void createLLVMFuncOpIfNotExist(mlir::ConversionPatternRewriter &rewriter, |
| 5412 | mlir::SymbolTableCollection &symbolTables, |
| 5413 | mlir::Operation *srcOp, llvm::StringRef fnName, |
| 5414 | mlir::Type fnTy, mlir::ArrayAttr argAttrs, |
| 5415 | mlir::ArrayAttr resAttrs) { |
| 5416 | mlir::ModuleOp modOp = srcOp->getParentOfType<mlir::ModuleOp>(); |
| 5417 | mlir::Operation *sourceSymbol = symbolTables.lookupSymbolIn( |
| 5418 | symbolTableOp: modOp, symbol: mlir::StringAttr::get(context: fnTy.getContext(), bytes: fnName)); |
| 5419 | if (!sourceSymbol) { |
| 5420 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 5421 | auto enclosingFnOp = srcOp->getParentOfType<mlir::LLVM::LLVMFuncOp>(); |
| 5422 | rewriter.setInsertionPoint(enclosingFnOp); |
| 5423 | auto fn = |
| 5424 | mlir::LLVM::LLVMFuncOp::create(builder&: rewriter, location: srcOp->getLoc(), name: fnName, type: fnTy); |
| 5425 | if (argAttrs) |
| 5426 | fn.setArgAttrsAttr(argAttrs); |
| 5427 | if (resAttrs) |
| 5428 | fn.setResAttrsAttr(resAttrs); |
| 5429 | // Update the symbol table. |
| 5430 | symbolTables.getSymbolTable(op: fn->getParentOp()).insert(symbol: fn); |
| 5431 | } |
| 5432 | } |
| 5433 | |
| 5434 | mlir::LogicalResult CIRToLLVMThrowOpLowering::matchAndRewrite( |
| 5435 | cir::ThrowOp op, OpAdaptor adaptor, |
| 5436 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5437 | mlir::Location loc = op.getLoc(); |
| 5438 | auto voidTy = mlir::LLVM::LLVMVoidType::get(ctx: getContext()); |
| 5439 | |
| 5440 | // TODO(cir): set the runtime calling convention on the runtime calls below. |
| 5441 | assert(!cir::MissingFeatures::opFuncCallingConv()); |
| 5442 | |
| 5443 | if (op.rethrows()) { |
| 5444 | auto funcTy = mlir::LLVM::LLVMFunctionType::get(result: voidTy, arguments: {}); |
| 5445 | |
| 5446 | // Get or create `declare void @__cxa_rethrow()` |
| 5447 | const llvm::StringRef functionName = "__cxa_rethrow" ; |
| 5448 | createLLVMFuncOpIfNotExist(rewriter, symbolTables, srcOp: op, fnName: functionName, |
| 5449 | fnTy: funcTy); |
| 5450 | |
| 5451 | auto cxaRethrow = mlir::LLVM::CallOp::create( |
| 5452 | builder&: rewriter, location: loc, results: mlir::TypeRange{}, callee: functionName); |
| 5453 | |
| 5454 | rewriter.replaceOp(op, newOp: cxaRethrow); |
| 5455 | return mlir::success(); |
| 5456 | } |
| 5457 | |
| 5458 | auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5459 | auto fnTy = mlir::LLVM::LLVMFunctionType::get( |
| 5460 | result: voidTy, arguments: {llvmPtrTy, llvmPtrTy, llvmPtrTy}); |
| 5461 | |
| 5462 | // Get or create `declare void @__cxa_throw(ptr, ptr, ptr)` |
| 5463 | const llvm::StringRef fnName = "__cxa_throw" ; |
| 5464 | createLLVMFuncOpIfNotExist(rewriter, symbolTables, srcOp: op, fnName, fnTy); |
| 5465 | |
| 5466 | mlir::Value typeInfo = mlir::LLVM::AddressOfOp::create( |
| 5467 | builder&: rewriter, location: loc, res: mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()), |
| 5468 | global_name: adaptor.getTypeInfoAttr()); |
| 5469 | |
| 5470 | mlir::Value dtor; |
| 5471 | if (op.getDtor()) { |
| 5472 | dtor = mlir::LLVM::AddressOfOp::create(builder&: rewriter, location: loc, res: llvmPtrTy, |
| 5473 | global_name: adaptor.getDtorAttr()); |
| 5474 | } else { |
| 5475 | dtor = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: llvmPtrTy); |
| 5476 | } |
| 5477 | |
| 5478 | auto cxaThrowCall = mlir::LLVM::CallOp::create( |
| 5479 | builder&: rewriter, location: loc, results: mlir::TypeRange{}, callee: fnName, |
| 5480 | args: mlir::ValueRange{adaptor.getExceptionPtr(), typeInfo, dtor}); |
| 5481 | |
| 5482 | rewriter.replaceOp(op, newOp: cxaThrowCall); |
| 5483 | return mlir::success(); |
| 5484 | } |
| 5485 | |
| 5486 | mlir::LogicalResult CIRToLLVMAllocExceptionOpLowering::matchAndRewrite( |
| 5487 | cir::AllocExceptionOp op, OpAdaptor adaptor, |
| 5488 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5489 | // Get or create `declare ptr @__cxa_allocate_exception(i64)` |
| 5490 | StringRef fnName = "__cxa_allocate_exception" ; |
| 5491 | auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5492 | auto int64Ty = mlir::IntegerType::get(context: rewriter.getContext(), width: 64); |
| 5493 | auto fnTy = mlir::LLVM::LLVMFunctionType::get(result: llvmPtrTy, arguments: {int64Ty}); |
| 5494 | |
| 5495 | createLLVMFuncOpIfNotExist(rewriter, symbolTables, srcOp: op, fnName, fnTy); |
| 5496 | auto exceptionSize = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: op.getLoc(), |
| 5497 | value: adaptor.getSizeAttr()); |
| 5498 | |
| 5499 | auto allocaExceptionCall = mlir::LLVM::CallOp::create( |
| 5500 | builder&: rewriter, location: op.getLoc(), results: mlir::TypeRange{llvmPtrTy}, callee: fnName, |
| 5501 | args: mlir::ValueRange{exceptionSize}); |
| 5502 | |
| 5503 | rewriter.replaceOp(op, newOp: allocaExceptionCall); |
| 5504 | return mlir::success(); |
| 5505 | } |
| 5506 | |
| 5507 | static mlir::LLVM::LLVMStructType |
| 5508 | getLLVMLandingPadStructTy(mlir::ConversionPatternRewriter &rewriter) { |
| 5509 | // Create the landing pad type: struct { ptr, i32 } |
| 5510 | mlir::MLIRContext *ctx = rewriter.getContext(); |
| 5511 | auto llvmPtr = mlir::LLVM::LLVMPointerType::get(context: ctx); |
| 5512 | llvm::SmallVector<mlir::Type> structFields = {llvmPtr, rewriter.getI32Type()}; |
| 5513 | return mlir::LLVM::LLVMStructType::getLiteral(context: ctx, types: structFields); |
| 5514 | } |
| 5515 | |
| 5516 | mlir::LogicalResult CIRToLLVMEhInflightOpLowering::matchAndRewrite( |
| 5517 | cir::EhInflightOp op, OpAdaptor adaptor, |
| 5518 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5519 | auto llvmFn = op->getParentOfType<mlir::LLVM::LLVMFuncOp>(); |
| 5520 | assert(llvmFn && "expected LLVM function parent" ); |
| 5521 | mlir::Block *entryBlock = &llvmFn.getRegion().front(); |
| 5522 | assert(entryBlock->isEntryBlock()); |
| 5523 | |
| 5524 | mlir::ArrayAttr catchListAttr = op.getCatchTypeListAttr(); |
| 5525 | mlir::ArrayAttr filterListAttr = op.getFilterTypeListAttr(); |
| 5526 | mlir::SmallVector<mlir::Value> landingPadClauses; |
| 5527 | |
| 5528 | auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5529 | mlir::Location loc = op.getLoc(); |
| 5530 | |
| 5531 | // %landingpad = landingpad { ptr, i32 } |
| 5532 | // Note that since llvm.landingpad has to be the first operation on the |
| 5533 | // block, any needed value for its operands has to be added somewhere else. |
| 5534 | if (catchListAttr) { |
| 5535 | // catch ptr @_ZTIi |
| 5536 | // catch ptr @_ZTIPKc |
| 5537 | for (mlir::Attribute catchAttr : catchListAttr) { |
| 5538 | auto symAttr = cast<mlir::FlatSymbolRefAttr>(Val&: catchAttr); |
| 5539 | // Generate `llvm.mlir.addressof` for each symbol, and place those |
| 5540 | // operations in the LLVM function entry basic block. |
| 5541 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 5542 | rewriter.setInsertionPointToStart(entryBlock); |
| 5543 | mlir::Value addrOp = mlir::LLVM::AddressOfOp::create( |
| 5544 | builder&: rewriter, location: loc, res: llvmPtrTy, global_name: symAttr.getValue()); |
| 5545 | landingPadClauses.push_back(Elt: addrOp); |
| 5546 | } |
| 5547 | } |
| 5548 | |
| 5549 | // Emit a catch-all clause (catch ptr null) when: |
| 5550 | // - The catch_all attribute is set (typed catches + catch-all), or |
| 5551 | // - No typed catches, no cleanup, and no filter (legacy pure catch-all) |
| 5552 | if (op.getCatchAll() || |
| 5553 | (!catchListAttr && !op.getCleanup() && !filterListAttr)) { |
| 5554 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 5555 | rewriter.setInsertionPointToStart(entryBlock); |
| 5556 | mlir::Value nullOp = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: llvmPtrTy); |
| 5557 | landingPadClauses.push_back(Elt: nullOp); |
| 5558 | } |
| 5559 | |
| 5560 | if (filterListAttr) { |
| 5561 | // Filter clauses are array-typed landingpad operands: |
| 5562 | // filter [N x ptr] [ptr @_ZTIi, ...] |
| 5563 | // An empty list is throw() and lowers to [0 x ptr] zeroinitializer. |
| 5564 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 5565 | rewriter.setInsertionPointToStart(entryBlock); |
| 5566 | auto filterArrayTy = |
| 5567 | mlir::LLVM::LLVMArrayType::get(elementType: llvmPtrTy, numElements: filterListAttr.size()); |
| 5568 | mlir::Value filterVal; |
| 5569 | if (filterListAttr.empty()) { |
| 5570 | filterVal = mlir::LLVM::ZeroOp::create(builder&: rewriter, location: loc, res: filterArrayTy); |
| 5571 | } else { |
| 5572 | filterVal = mlir::LLVM::UndefOp::create(builder&: rewriter, location: loc, res: filterArrayTy); |
| 5573 | for (auto [i, filterAttr] : llvm::enumerate(First&: filterListAttr)) { |
| 5574 | auto symAttr = mlir::cast<mlir::FlatSymbolRefAttr>(Val: filterAttr); |
| 5575 | mlir::Value addrOp = mlir::LLVM::AddressOfOp::create( |
| 5576 | builder&: rewriter, location: loc, res: llvmPtrTy, global_name: symAttr.getValue()); |
| 5577 | filterVal = mlir::LLVM::InsertValueOp::create( |
| 5578 | builder&: rewriter, location: loc, container: filterVal, value: addrOp, |
| 5579 | position: llvm::ArrayRef<int64_t>{static_cast<int64_t>(i)}); |
| 5580 | } |
| 5581 | } |
| 5582 | landingPadClauses.push_back(Elt: filterVal); |
| 5583 | } |
| 5584 | |
| 5585 | // %slot = extractvalue { ptr, i32 } %x, 0 |
| 5586 | // %selector = extractvalue { ptr, i32 } %x, 1 |
| 5587 | mlir::LLVM::LLVMStructType llvmLandingPadStructTy = |
| 5588 | getLLVMLandingPadStructTy(rewriter); |
| 5589 | auto landingPadOp = mlir::LLVM::LandingpadOp::create( |
| 5590 | builder&: rewriter, location: loc, resultType: llvmLandingPadStructTy, operands: landingPadClauses); |
| 5591 | |
| 5592 | // The LLVM cleanup flag is only needed when there is no catch-all handler, |
| 5593 | // since catch-all (catch ptr null) already ensures the personality function |
| 5594 | // enters the landing pad for all exception types. |
| 5595 | if (op.getCleanup() && !op.getCatchAll()) |
| 5596 | landingPadOp.setCleanup(true); |
| 5597 | |
| 5598 | mlir::Value slot = |
| 5599 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: landingPadOp, position: 0); |
| 5600 | mlir::Value selector = |
| 5601 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: landingPadOp, position: 1); |
| 5602 | rewriter.replaceOp(op, newValues: mlir::ValueRange{slot, selector}); |
| 5603 | |
| 5604 | return mlir::success(); |
| 5605 | } |
| 5606 | |
| 5607 | mlir::LogicalResult CIRToLLVMResumeFlatOpLowering::matchAndRewrite( |
| 5608 | cir::ResumeFlatOp op, OpAdaptor adaptor, |
| 5609 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5610 | // %lpad.val = insertvalue { ptr, i32 } poison, ptr %exception_ptr, 0 |
| 5611 | // %lpad.val2 = insertvalue { ptr, i32 } %lpad.val, i32 %selector, 1 |
| 5612 | // resume { ptr, i32 } %lpad.val2 |
| 5613 | mlir::Type llvmLandingPadStructTy = getLLVMLandingPadStructTy(rewriter); |
| 5614 | mlir::Value poison = mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op.getLoc(), |
| 5615 | res: llvmLandingPadStructTy); |
| 5616 | |
| 5617 | SmallVector<int64_t> slotIdx = {0}; |
| 5618 | mlir::Value slot = mlir::LLVM::InsertValueOp::create( |
| 5619 | builder&: rewriter, location: op.getLoc(), container: poison, value: adaptor.getExceptionPtr(), position: slotIdx); |
| 5620 | |
| 5621 | SmallVector<int64_t> selectorIdx = {1}; |
| 5622 | mlir::Value selector = mlir::LLVM::InsertValueOp::create( |
| 5623 | builder&: rewriter, location: op.getLoc(), container: slot, value: adaptor.getTypeId(), position: selectorIdx); |
| 5624 | |
| 5625 | rewriter.replaceOpWithNewOp<mlir::LLVM::ResumeOp>(op, args&: selector); |
| 5626 | return mlir::success(); |
| 5627 | } |
| 5628 | |
| 5629 | mlir::LogicalResult CIRToLLVMEhTypeIdOpLowering::matchAndRewrite( |
| 5630 | cir::EhTypeIdOp op, OpAdaptor adaptor, |
| 5631 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5632 | mlir::Value addrOp = mlir::LLVM::AddressOfOp::create( |
| 5633 | builder&: rewriter, location: op.getLoc(), |
| 5634 | res: mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()), |
| 5635 | global_name: op.getTypeSymAttr()); |
| 5636 | rewriter.replaceOpWithNewOp<mlir::LLVM::EhTypeidForOp>( |
| 5637 | op, args: rewriter.getI32Type(), args&: addrOp); |
| 5638 | return mlir::success(); |
| 5639 | } |
| 5640 | |
| 5641 | mlir::LogicalResult CIRToLLVMEhSetjmpOpLowering::matchAndRewrite( |
| 5642 | cir::EhSetjmpOp op, OpAdaptor adaptor, |
| 5643 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5644 | mlir::Type returnType = typeConverter->convertType(t: op.getType()); |
| 5645 | mlir::LLVM::CallIntrinsicOp newOp = |
| 5646 | createCallLLVMIntrinsicOp(rewriter, loc: op.getLoc(), intrinsicName: "llvm.eh.sjlj.setjmp" , |
| 5647 | resultTy: returnType, operands: adaptor.getEnv()); |
| 5648 | rewriter.replaceOp(op, newOp); |
| 5649 | return mlir::success(); |
| 5650 | } |
| 5651 | |
| 5652 | mlir::LogicalResult CIRToLLVMEhLongjmpOpLowering::matchAndRewrite( |
| 5653 | cir::EhLongjmpOp op, OpAdaptor adaptor, |
| 5654 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5655 | replaceOpWithCallLLVMIntrinsicOp(rewriter, op, intrinsicName: "llvm.eh.sjlj.longjmp" , |
| 5656 | /*resultTy=*/{}, operands: adaptor.getOperands()); |
| 5657 | return mlir::success(); |
| 5658 | } |
| 5659 | |
| 5660 | mlir::LogicalResult CIRToLLVMTrapOpLowering::matchAndRewrite( |
| 5661 | cir::TrapOp op, OpAdaptor adaptor, |
| 5662 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5663 | mlir::Location loc = op->getLoc(); |
| 5664 | rewriter.eraseOp(op); |
| 5665 | |
| 5666 | mlir::LLVM::Trap::create(builder&: rewriter, location: loc); |
| 5667 | |
| 5668 | // Note that the call to llvm.trap is not a terminator in LLVM dialect. |
| 5669 | // So we must emit an additional llvm.unreachable to terminate the current |
| 5670 | // block. |
| 5671 | mlir::LLVM::UnreachableOp::create(builder&: rewriter, location: loc); |
| 5672 | |
| 5673 | return mlir::success(); |
| 5674 | } |
| 5675 | |
| 5676 | static mlir::Value |
| 5677 | getValueForVTableSymbol(mlir::Operation *op, |
| 5678 | mlir::ConversionPatternRewriter &rewriter, |
| 5679 | mlir::SymbolTableCollection &symbolTables, |
| 5680 | const mlir::TypeConverter *converter, |
| 5681 | const CIRPreConversionTypeMap &preConversionTypes, |
| 5682 | mlir::FlatSymbolRefAttr nameAttr, mlir::Type &eltType) { |
| 5683 | auto module = op->getParentOfType<mlir::ModuleOp>(); |
| 5684 | mlir::Operation *symbol = symbolTables.lookupSymbolIn(symbolTableOp: module, name: nameAttr); |
| 5685 | if (auto llvmSymbol = mlir::dyn_cast<mlir::LLVM::GlobalOp>(Val: symbol)) { |
| 5686 | if (mlir::Type preType = preConversionTypes.lookup(Key: nameAttr.getValue())) { |
| 5687 | eltType = converter->convertType(t: preType); |
| 5688 | if (!eltType) { |
| 5689 | emitTypeConversionError(op, type: preType) |
| 5690 | << " for vtable symbol " << nameAttr; |
| 5691 | return {}; |
| 5692 | } |
| 5693 | } else { |
| 5694 | eltType = llvmSymbol.getType(); |
| 5695 | } |
| 5696 | } else if (auto cirSymbol = mlir::dyn_cast<cir::GlobalOp>(Val: symbol)) { |
| 5697 | eltType = converter->convertType(t: cirSymbol.getSymType()); |
| 5698 | if (!eltType) { |
| 5699 | emitTypeConversionError(op, type: cirSymbol.getSymType()) |
| 5700 | << " for vtable symbol " << nameAttr; |
| 5701 | return {}; |
| 5702 | } |
| 5703 | } else { |
| 5704 | op->emitError() << "unexpected symbol type for " << symbol; |
| 5705 | return {}; |
| 5706 | } |
| 5707 | |
| 5708 | return mlir::LLVM::AddressOfOp::create( |
| 5709 | builder&: rewriter, location: op->getLoc(), |
| 5710 | res: mlir::LLVM::LLVMPointerType::get(context: op->getContext()), global_name: nameAttr.getValue()); |
| 5711 | } |
| 5712 | |
| 5713 | mlir::LogicalResult CIRToLLVMVTableAddrPointOpLowering::matchAndRewrite( |
| 5714 | cir::VTableAddrPointOp op, OpAdaptor adaptor, |
| 5715 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5716 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 5717 | if (usesRelativeVTables(mod: op->getParentOfType<mlir::ModuleOp>())) |
| 5718 | return op.emitError(message: "NYI: address point of a relative layout vtable" ); |
| 5719 | |
| 5720 | const mlir::TypeConverter *converter = getTypeConverter(); |
| 5721 | mlir::Type targetType = converter->convertType(t: op.getType()); |
| 5722 | llvm::SmallVector<mlir::LLVM::GEPArg> offsets; |
| 5723 | mlir::Type eltType; |
| 5724 | mlir::Value symAddr = |
| 5725 | getValueForVTableSymbol(op, rewriter, symbolTables, converter, |
| 5726 | preConversionTypes, nameAttr: op.getNameAttr(), eltType); |
| 5727 | if (!symAddr) |
| 5728 | return mlir::failure(); |
| 5729 | |
| 5730 | offsets = llvm::SmallVector<mlir::LLVM::GEPArg>{ |
| 5731 | 0, op.getAddressPointAttr().getIndex(), |
| 5732 | op.getAddressPointAttr().getOffset()}; |
| 5733 | |
| 5734 | assert(eltType && "Shouldn't ever be missing an eltType here" ); |
| 5735 | mlir::LLVM::GEPNoWrapFlags inboundsNuw = |
| 5736 | mlir::LLVM::GEPNoWrapFlags::inbounds | mlir::LLVM::GEPNoWrapFlags::nuw; |
| 5737 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>(op, args&: targetType, args&: eltType, |
| 5738 | args&: symAddr, args&: offsets, args&: inboundsNuw); |
| 5739 | return mlir::success(); |
| 5740 | } |
| 5741 | |
| 5742 | mlir::LogicalResult CIRToLLVMVTableGetVPtrOpLowering::matchAndRewrite( |
| 5743 | cir::VTableGetVPtrOp op, OpAdaptor adaptor, |
| 5744 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5745 | // cir.vtable.get_vptr is equivalent to a bitcast from the source object |
| 5746 | // pointer to the vptr type. Since the LLVM dialect uses opaque pointers |
| 5747 | // we can just replace uses of this operation with the original pointer. |
| 5748 | mlir::Value srcVal = adaptor.getSrc(); |
| 5749 | rewriter.replaceOp(op, newValues: srcVal); |
| 5750 | return mlir::success(); |
| 5751 | } |
| 5752 | |
| 5753 | mlir::LogicalResult CIRToLLVMVTableGetVirtualFnOpLowering::matchAndRewrite( |
| 5754 | cir::VTableGetVirtualFnOp op, OpAdaptor adaptor, |
| 5755 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5756 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 5757 | if (usesRelativeVTables(mod: op->getParentOfType<mlir::ModuleOp>())) |
| 5758 | return op.emitError( |
| 5759 | message: "NYI: virtual function lookup in a relative layout vtable" ); |
| 5760 | |
| 5761 | mlir::Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 5762 | if (!resultType) |
| 5763 | return emitTypeConversionError(op, type: op.getType()); |
| 5764 | |
| 5765 | // The entry is the function pointer itself: load it from its slot. |
| 5766 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5767 | mlir::Value slot = mlir::LLVM::GEPOp::create( |
| 5768 | builder&: rewriter, location: op.getLoc(), resultType: ptrTy, elementType: ptrTy, basePtr: adaptor.getVptr(), |
| 5769 | indices: llvm::ArrayRef<mlir::LLVM::GEPArg>{static_cast<int32_t>(op.getIndex())}, |
| 5770 | noWrapFlags: mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 5771 | rewriter.replaceOpWithNewOp<mlir::LLVM::LoadOp>( |
| 5772 | op, args&: resultType, args&: slot, args: dataLayout.getTypeABIAlignment(t: ptrTy), |
| 5773 | /*isVolatile=*/args: false, /*isNonTemporal=*/args: false, |
| 5774 | /*isInvariant=*/args: op.getInvariant(), /*isInvariantGroup=*/args: false); |
| 5775 | return mlir::success(); |
| 5776 | } |
| 5777 | |
| 5778 | mlir::LogicalResult |
| 5779 | CIRToLLVMVTableGetVirtualFnByOffsetOpLowering::matchAndRewrite( |
| 5780 | cir::VTableGetVirtualFnByOffsetOp op, OpAdaptor adaptor, |
| 5781 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5782 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 5783 | if (usesRelativeVTables(mod: op->getParentOfType<mlir::ModuleOp>())) |
| 5784 | return op.emitError( |
| 5785 | message: "NYI: virtual function lookup by offset in a relative layout vtable" ); |
| 5786 | |
| 5787 | mlir::Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 5788 | if (!resultType) |
| 5789 | return emitTypeConversionError(op, type: op.getType()); |
| 5790 | |
| 5791 | // The entry is the function pointer itself: load it from its slot, addressed |
| 5792 | // by its byte offset. |
| 5793 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5794 | mlir::Value slot = mlir::LLVM::GEPOp::create( |
| 5795 | builder&: rewriter, location: op.getLoc(), resultType: ptrTy, elementType: rewriter.getI8Type(), basePtr: adaptor.getVptr(), |
| 5796 | indices: llvm::ArrayRef<mlir::LLVM::GEPArg>{adaptor.getOffset()}); |
| 5797 | rewriter.replaceOpWithNewOp<mlir::LLVM::LoadOp>( |
| 5798 | op, args&: resultType, args&: slot, args: dataLayout.getTypeABIAlignment(t: ptrTy)); |
| 5799 | return mlir::success(); |
| 5800 | } |
| 5801 | |
| 5802 | mlir::LogicalResult CIRToLLVMVTableGetOffsetOpLowering::matchAndRewrite( |
| 5803 | cir::VTableGetOffsetOp op, OpAdaptor adaptor, |
| 5804 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5805 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 5806 | if (usesRelativeVTables(mod: op->getParentOfType<mlir::ModuleOp>())) |
| 5807 | return op.emitError(message: "NYI: offset lookup in a relative layout vtable" ); |
| 5808 | |
| 5809 | mlir::Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 5810 | if (!resultType) |
| 5811 | return emitTypeConversionError(op, type: op.getType()); |
| 5812 | |
| 5813 | // The entry is the offset itself: load it from its slot, addressed by its |
| 5814 | // byte offset. |
| 5815 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5816 | int64_t byteOffset = |
| 5817 | op.getIndex() * cir::getVTableEntrySize( |
| 5818 | mod: op->getParentOfType<mlir::ModuleOp>(), dataLayout); |
| 5819 | mlir::Value byteOffsetVal = mlir::LLVM::ConstantOp::create( |
| 5820 | builder&: rewriter, location: op.getLoc(), res: rewriter.getI64Type(), |
| 5821 | value: rewriter.getI64IntegerAttr(value: byteOffset)); |
| 5822 | mlir::Value slot = mlir::LLVM::GEPOp::create( |
| 5823 | builder&: rewriter, location: op.getLoc(), resultType: ptrTy, elementType: rewriter.getI8Type(), basePtr: adaptor.getVptr(), |
| 5824 | indices: llvm::ArrayRef<mlir::LLVM::GEPArg>{byteOffsetVal}); |
| 5825 | rewriter.replaceOpWithNewOp<mlir::LLVM::LoadOp>( |
| 5826 | op, args&: resultType, args&: slot, args: dataLayout.getTypeABIAlignment(t: ptrTy)); |
| 5827 | return mlir::success(); |
| 5828 | } |
| 5829 | |
| 5830 | mlir::LogicalResult CIRToLLVMVTableGetEntryOpLowering::matchAndRewrite( |
| 5831 | cir::VTableGetEntryOp op, OpAdaptor adaptor, |
| 5832 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5833 | assert(!cir::MissingFeatures::vtableRelativeLayout()); |
| 5834 | if (usesRelativeVTables(mod: op->getParentOfType<mlir::ModuleOp>())) |
| 5835 | return op.emitError( |
| 5836 | message: "NYI: pointer entry lookup in a relative layout vtable" ); |
| 5837 | |
| 5838 | mlir::Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 5839 | if (!resultType) |
| 5840 | return emitTypeConversionError(op, type: op.getType()); |
| 5841 | |
| 5842 | // The entry is the pointer itself: load it from its slot. |
| 5843 | auto ptrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5844 | mlir::Value index = mlir::LLVM::ConstantOp::create( |
| 5845 | builder&: rewriter, location: op.getLoc(), res: rewriter.getI64Type(), |
| 5846 | value: rewriter.getI64IntegerAttr(value: op.getIndex())); |
| 5847 | mlir::Value slot = mlir::LLVM::GEPOp::create( |
| 5848 | builder&: rewriter, location: op.getLoc(), resultType: ptrTy, elementType: ptrTy, basePtr: adaptor.getVptr(), |
| 5849 | indices: llvm::ArrayRef<mlir::LLVM::GEPArg>{index}, |
| 5850 | noWrapFlags: mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 5851 | rewriter.replaceOpWithNewOp<mlir::LLVM::LoadOp>( |
| 5852 | op, args&: resultType, args&: slot, args: dataLayout.getTypeABIAlignment(t: ptrTy)); |
| 5853 | return mlir::success(); |
| 5854 | } |
| 5855 | |
| 5856 | mlir::LogicalResult CIRToLLVMVTTAddrPointOpLowering::matchAndRewrite( |
| 5857 | cir::VTTAddrPointOp op, OpAdaptor adaptor, |
| 5858 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5859 | const mlir::Type resultType = getTypeConverter()->convertType(t: op.getType()); |
| 5860 | if (!resultType) |
| 5861 | return emitTypeConversionError(op, type: op.getType()); |
| 5862 | llvm::SmallVector<mlir::LLVM::GEPArg> offsets; |
| 5863 | mlir::Type eltType; |
| 5864 | mlir::Value llvmAddr = adaptor.getSymAddr(); |
| 5865 | |
| 5866 | if (op.getSymAddr()) { |
| 5867 | if (op.getOffset() == 0) { |
| 5868 | rewriter.replaceOp(op, newValues: {llvmAddr}); |
| 5869 | return mlir::success(); |
| 5870 | } |
| 5871 | |
| 5872 | offsets.push_back(Elt: adaptor.getOffset()); |
| 5873 | eltType = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 5874 | } else { |
| 5875 | llvmAddr = |
| 5876 | getValueForVTableSymbol(op, rewriter, symbolTables, converter: getTypeConverter(), |
| 5877 | preConversionTypes, nameAttr: op.getNameAttr(), eltType); |
| 5878 | if (!llvmAddr) |
| 5879 | return mlir::failure(); |
| 5880 | assert(eltType && "Shouldn't ever be missing an eltType here" ); |
| 5881 | offsets.push_back(Elt: 0); |
| 5882 | offsets.push_back(Elt: adaptor.getOffset()); |
| 5883 | } |
| 5884 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 5885 | op, args: resultType, args&: eltType, args&: llvmAddr, args&: offsets, |
| 5886 | args: mlir::LLVM::GEPNoWrapFlags::inbounds); |
| 5887 | return mlir::success(); |
| 5888 | } |
| 5889 | |
| 5890 | mlir::LogicalResult CIRToLLVMVecCreateOpLowering::matchAndRewrite( |
| 5891 | cir::VecCreateOp op, OpAdaptor adaptor, |
| 5892 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5893 | // Start with an 'undef' value for the vector. Then 'insertelement' for |
| 5894 | // each of the vector elements. |
| 5895 | const cir::VectorType vecTy = op.getType(); |
| 5896 | const mlir::Type llvmTy = typeConverter->convertType(t: vecTy); |
| 5897 | const mlir::Location loc = op.getLoc(); |
| 5898 | mlir::Value result = mlir::LLVM::PoisonOp::create(builder&: rewriter, location: loc, res: llvmTy); |
| 5899 | const uint64_t vectorSize = vecTy.getFixedSize(); |
| 5900 | assert(vectorSize == op.getElements().size() && |
| 5901 | "cir.vec.create op count doesn't match vector type elements count" ); |
| 5902 | |
| 5903 | for (uint64_t i = 0; i < vectorSize; ++i) { |
| 5904 | const mlir::Value indexValue = |
| 5905 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI64Type(), value: i); |
| 5906 | result = mlir::LLVM::InsertElementOp::create( |
| 5907 | builder&: rewriter, location: loc, vector: result, value: adaptor.getElements()[i], position: indexValue); |
| 5908 | } |
| 5909 | |
| 5910 | rewriter.replaceOp(op, newValues: result); |
| 5911 | return mlir::success(); |
| 5912 | } |
| 5913 | |
| 5914 | mlir::LogicalResult CIRToLLVMVecCmpOpLowering::matchAndRewrite( |
| 5915 | cir::VecCmpOp op, OpAdaptor adaptor, |
| 5916 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5917 | mlir::Type elementType = op.getLhs().getType().getElementType(); |
| 5918 | mlir::Value bitResult; |
| 5919 | if (auto intType = mlir::dyn_cast<cir::IntType>(Val&: elementType)) { |
| 5920 | bitResult = mlir::LLVM::ICmpOp::create( |
| 5921 | builder&: rewriter, location: op.getLoc(), |
| 5922 | predicate: convertCmpKindToICmpPredicate(kind: op.getKind(), isSigned: intType.isSigned()), |
| 5923 | lhs: adaptor.getLhs(), rhs: adaptor.getRhs()); |
| 5924 | } else if (auto boolType = mlir::dyn_cast<cir::BoolType>(Val&: elementType)) { |
| 5925 | bitResult = mlir::LLVM::ICmpOp::create( |
| 5926 | builder&: rewriter, location: op.getLoc(), |
| 5927 | predicate: convertCmpKindToICmpPredicate(kind: op.getKind(), /*isSigned=*/false), |
| 5928 | lhs: adaptor.getLhs(), rhs: adaptor.getRhs()); |
| 5929 | } else { |
| 5930 | return op.emitError() << "unsupported type for VecCmpOp: " << elementType; |
| 5931 | } |
| 5932 | |
| 5933 | // LLVM IR vector comparison returns a vector of i1. This one-bit vector |
| 5934 | // must be sign-extended to the correct result type, unless a vector of i1 is |
| 5935 | // the type we need. |
| 5936 | mlir::Type vecElemTy = op.getType().getElementType(); |
| 5937 | if (isa<cir::IntType>(Val: vecElemTy) && |
| 5938 | cast<cir::IntType>(Val&: vecElemTy).getWidth() > 1) |
| 5939 | rewriter.replaceOpWithNewOp<mlir::LLVM::SExtOp>( |
| 5940 | op, args: typeConverter->convertType(t: op.getType()), args&: bitResult); |
| 5941 | else |
| 5942 | rewriter.replaceOp(op, newValues: bitResult); |
| 5943 | return mlir::success(); |
| 5944 | } |
| 5945 | |
| 5946 | mlir::LogicalResult CIRToLLVMVecFCmpOpLowering::matchAndRewrite( |
| 5947 | cir::VecFCmpOp op, OpAdaptor adaptor, |
| 5948 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5949 | mlir::Type elementType = op.getLhs().getType().getElementType(); |
| 5950 | if (!mlir::isa<cir::FPTypeInterface>(Val: elementType)) |
| 5951 | return op.emitError() << "unsupported type for VecFCmpOp: " << elementType; |
| 5952 | |
| 5953 | mlir::Value bitResult; |
| 5954 | if (cir::FenvAttr fenv = op.getFenvAttr()) { |
| 5955 | llvm::ElementCount size = op.getLhs().getType().getSize(); |
| 5956 | auto i1VecTy = mlir::VectorType::get( |
| 5957 | shape: size.getKnownMinValue(), elementType: rewriter.getI1Type(), scalableDims: {size.isScalable()}); |
| 5958 | mlir::LLVM::CallIntrinsicOp CFCmpCall = createConstrainedFCmpCall( |
| 5959 | rewriter, loc: op.getLoc(), lhs: adaptor.getLhs(), rhs: adaptor.getRhs(), kind: op.getKind(), |
| 5960 | isSignaling: op.getSignaling(), fenv, llvmResTy: i1VecTy); |
| 5961 | bitResult = CFCmpCall.getResult(i: 0); |
| 5962 | } else { |
| 5963 | bitResult = mlir::LLVM::FCmpOp::create( |
| 5964 | builder&: rewriter, location: op.getLoc(), predicate: convertCmpKindToFCmpPredicate(kind: op.getKind()), |
| 5965 | lhs: adaptor.getLhs(), rhs: adaptor.getRhs()); |
| 5966 | } |
| 5967 | |
| 5968 | // LLVM IR vector comparison returns a vector of i1. This one-bit vector |
| 5969 | // must be sign-extended to the correct result type, unless a vector of i1 is |
| 5970 | // the type we need. |
| 5971 | mlir::Type vecElemTy = op.getType().getElementType(); |
| 5972 | if (isa<cir::IntType>(Val: vecElemTy) && |
| 5973 | cast<cir::IntType>(Val&: vecElemTy).getWidth() > 1) |
| 5974 | rewriter.replaceOpWithNewOp<mlir::LLVM::SExtOp>( |
| 5975 | op, args: typeConverter->convertType(t: op.getType()), args&: bitResult); |
| 5976 | else |
| 5977 | rewriter.replaceOp(op, newValues: bitResult); |
| 5978 | return mlir::success(); |
| 5979 | } |
| 5980 | |
| 5981 | mlir::LogicalResult CIRToLLVMVecSplatOpLowering::matchAndRewrite( |
| 5982 | cir::VecSplatOp op, OpAdaptor adaptor, |
| 5983 | mlir::ConversionPatternRewriter &rewriter) const { |
| 5984 | // Vector splat can be implemented with an `insertelement` and a |
| 5985 | // `shufflevector`, which is better than an `insertelement` for each |
| 5986 | // element in the vector. Start with an undef vector. Insert the value into |
| 5987 | // the first element. Then use a `shufflevector` with a mask of all 0 to |
| 5988 | // fill out the entire vector with that value. |
| 5989 | mlir::Type llvmTy = typeConverter->convertType(t: op.getType()); |
| 5990 | mlir::Location loc = op.getLoc(); |
| 5991 | mlir::Value poison = mlir::LLVM::PoisonOp::create(builder&: rewriter, location: loc, res: llvmTy); |
| 5992 | |
| 5993 | mlir::Value elementValue = adaptor.getValue(); |
| 5994 | if (elementValue.getDefiningOp<mlir::LLVM::PoisonOp>()) { |
| 5995 | // If the splat value is poison, then we can just use poison value |
| 5996 | // for the entire vector. |
| 5997 | rewriter.replaceOp(op, newValues: poison); |
| 5998 | return mlir::success(); |
| 5999 | } |
| 6000 | |
| 6001 | if (auto constValue = elementValue.getDefiningOp<mlir::LLVM::ConstantOp>()) { |
| 6002 | if (auto intAttr = dyn_cast<mlir::IntegerAttr>(Val: constValue.getValue())) { |
| 6003 | mlir::DenseIntElementsAttr denseVec = mlir::DenseIntElementsAttr::get( |
| 6004 | type: mlir::cast<mlir::ShapedType>(Val&: llvmTy), arg: intAttr.getValue()); |
| 6005 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( |
| 6006 | op, args: denseVec.getType(), args&: denseVec); |
| 6007 | return mlir::success(); |
| 6008 | } |
| 6009 | |
| 6010 | if (auto fpAttr = dyn_cast<mlir::FloatAttr>(Val: constValue.getValue())) { |
| 6011 | mlir::DenseFPElementsAttr denseVec = mlir::DenseFPElementsAttr::get( |
| 6012 | type: mlir::cast<mlir::ShapedType>(Val&: llvmTy), arg: fpAttr.getValue()); |
| 6013 | rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>( |
| 6014 | op, args: denseVec.getType(), args&: denseVec); |
| 6015 | return mlir::success(); |
| 6016 | } |
| 6017 | } |
| 6018 | |
| 6019 | mlir::Value indexValue = |
| 6020 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI64Type(), value: 0); |
| 6021 | mlir::Value oneElement = mlir::LLVM::InsertElementOp::create( |
| 6022 | builder&: rewriter, location: loc, vector: poison, value: elementValue, position: indexValue); |
| 6023 | SmallVector<int32_t> zeroValues( |
| 6024 | mlir::cast<mlir::VectorType>(Val&: llvmTy).getNumElements(), 0); |
| 6025 | rewriter.replaceOpWithNewOp<mlir::LLVM::ShuffleVectorOp>(op, args&: oneElement, |
| 6026 | args&: poison, args&: zeroValues); |
| 6027 | return mlir::success(); |
| 6028 | } |
| 6029 | |
| 6030 | mlir::LogicalResult CIRToLLVMVecShuffleOpLowering::matchAndRewrite( |
| 6031 | cir::VecShuffleOp op, OpAdaptor adaptor, |
| 6032 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6033 | // LLVM::ShuffleVectorOp takes an ArrayRef of int for the list of indices. |
| 6034 | // Convert the ClangIR ArrayAttr of IntAttr constants into a |
| 6035 | // SmallVector<int>. |
| 6036 | SmallVector<int, 8> indices; |
| 6037 | std::transform( |
| 6038 | first: op.getIndices().begin(), last: op.getIndices().end(), |
| 6039 | result: std::back_inserter(x&: indices), unary_op: [](mlir::Attribute intAttr) { |
| 6040 | return mlir::cast<cir::IntAttr>(Val&: intAttr).getValue().getSExtValue(); |
| 6041 | }); |
| 6042 | rewriter.replaceOpWithNewOp<mlir::LLVM::ShuffleVectorOp>( |
| 6043 | op, args: adaptor.getVec1(), args: adaptor.getVec2(), args&: indices); |
| 6044 | return mlir::success(); |
| 6045 | } |
| 6046 | |
| 6047 | mlir::LogicalResult CIRToLLVMVecShuffleDynamicOpLowering::matchAndRewrite( |
| 6048 | cir::VecShuffleDynamicOp op, OpAdaptor adaptor, |
| 6049 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6050 | // LLVM IR does not have an operation that corresponds to this form of |
| 6051 | // the built-in. |
| 6052 | // __builtin_shufflevector(V, I) |
| 6053 | // is implemented as this pseudocode, where the for loop is unrolled |
| 6054 | // and N is the number of elements: |
| 6055 | // |
| 6056 | // result = undef |
| 6057 | // maskbits = NextPowerOf2(N - 1) |
| 6058 | // masked = I & maskbits |
| 6059 | // for (i in 0 <= i < N) |
| 6060 | // result[i] = V[masked[i]] |
| 6061 | mlir::Location loc = op.getLoc(); |
| 6062 | mlir::Value input = adaptor.getVec(); |
| 6063 | mlir::Type llvmIndexVecType = |
| 6064 | getTypeConverter()->convertType(t: op.getIndices().getType()); |
| 6065 | mlir::Type llvmIndexType = getTypeConverter()->convertType( |
| 6066 | t: op.getIndices().getType().getElementType()); |
| 6067 | uint64_t numElements = op.getVec().getType().getFixedSize(); |
| 6068 | |
| 6069 | uint64_t maskBits = llvm::NextPowerOf2(A: numElements - 1) - 1; |
| 6070 | mlir::Value maskValue = mlir::LLVM::ConstantOp::create( |
| 6071 | builder&: rewriter, location: loc, res: llvmIndexType, |
| 6072 | value: rewriter.getIntegerAttr(type: llvmIndexType, value: maskBits)); |
| 6073 | mlir::Value maskVector = |
| 6074 | mlir::LLVM::UndefOp::create(builder&: rewriter, location: loc, res: llvmIndexVecType); |
| 6075 | |
| 6076 | for (uint64_t i = 0; i < numElements; ++i) { |
| 6077 | mlir::Value idxValue = |
| 6078 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI64Type(), value: i); |
| 6079 | maskVector = mlir::LLVM::InsertElementOp::create(builder&: rewriter, location: loc, vector: maskVector, |
| 6080 | value: maskValue, position: idxValue); |
| 6081 | } |
| 6082 | |
| 6083 | mlir::Value maskedIndices = mlir::LLVM::AndOp::create( |
| 6084 | builder&: rewriter, location: loc, res: llvmIndexVecType, lhs: adaptor.getIndices(), rhs: maskVector); |
| 6085 | mlir::Value result = mlir::LLVM::UndefOp::create( |
| 6086 | builder&: rewriter, location: loc, res: getTypeConverter()->convertType(t: op.getVec().getType())); |
| 6087 | for (uint64_t i = 0; i < numElements; ++i) { |
| 6088 | mlir::Value iValue = |
| 6089 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI64Type(), value: i); |
| 6090 | mlir::Value indexValue = mlir::LLVM::ExtractElementOp::create( |
| 6091 | builder&: rewriter, location: loc, vector: maskedIndices, position: iValue); |
| 6092 | mlir::Value valueAtIndex = |
| 6093 | mlir::LLVM::ExtractElementOp::create(builder&: rewriter, location: loc, vector: input, position: indexValue); |
| 6094 | result = mlir::LLVM::InsertElementOp::create(builder&: rewriter, location: loc, vector: result, |
| 6095 | value: valueAtIndex, position: iValue); |
| 6096 | } |
| 6097 | rewriter.replaceOp(op, newValues: result); |
| 6098 | return mlir::success(); |
| 6099 | } |
| 6100 | |
| 6101 | mlir::LogicalResult CIRToLLVMVecTernaryOpLowering::matchAndRewrite( |
| 6102 | cir::VecTernaryOp op, OpAdaptor adaptor, |
| 6103 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6104 | mlir::Value bitVec = adaptor.getCond(); |
| 6105 | |
| 6106 | if (!isa<cir::BoolType>(Val: op.getCond().getType().getElementType())) { |
| 6107 | // Convert `cond` into a vector of i1, then use that in a `select` op. |
| 6108 | bitVec = mlir::LLVM::ICmpOp::create( |
| 6109 | builder&: rewriter, location: op.getLoc(), predicate: mlir::LLVM::ICmpPredicate::ne, lhs: adaptor.getCond(), |
| 6110 | rhs: mlir::LLVM::ZeroOp::create( |
| 6111 | builder&: rewriter, location: op.getCond().getLoc(), |
| 6112 | res: typeConverter->convertType(t: op.getCond().getType()))); |
| 6113 | } |
| 6114 | |
| 6115 | rewriter.replaceOpWithNewOp<mlir::LLVM::SelectOp>( |
| 6116 | op, args&: bitVec, args: adaptor.getLhs(), args: adaptor.getRhs()); |
| 6117 | return mlir::success(); |
| 6118 | } |
| 6119 | |
| 6120 | /// Computes the column-major flat index (columnIdx * numRows) + rowIdx. |
| 6121 | static mlir::Value |
| 6122 | computeMatrixFlatIndex(mlir::ConversionPatternRewriter &rewriter, |
| 6123 | mlir::Location loc, cir::MatrixType matrixTy, |
| 6124 | mlir::Value rowIdx, mlir::Value columnIdx) { |
| 6125 | rowIdx = createIntCast(bld&: rewriter, src: rowIdx, dstTy: rewriter.getI64Type(), isSigned: false); |
| 6126 | columnIdx = createIntCast(bld&: rewriter, src: columnIdx, dstTy: rewriter.getI64Type(), isSigned: false); |
| 6127 | mlir::Type indexTy = rowIdx.getType(); |
| 6128 | mlir::Value numRows = mlir::LLVM::ConstantOp::create( |
| 6129 | builder&: rewriter, location: loc, res: indexTy, |
| 6130 | value: rewriter.getIntegerAttr(type: indexTy, value: matrixTy.getNumRows())); |
| 6131 | mlir::Value flatIndex = mlir::LLVM::MulOp::create( |
| 6132 | builder&: rewriter, location: loc, operands: {columnIdx, numRows}, /*properties=*/{}, |
| 6133 | /*discardableAttributes=*/{}); |
| 6134 | return mlir::LLVM::AddOp::create(builder&: rewriter, location: loc, lhs: flatIndex, rhs: rowIdx, |
| 6135 | /*overflowFlags=*/{}); |
| 6136 | } |
| 6137 | |
| 6138 | mlir::LogicalResult CIRToLLVMMatrixExtractOpLowering::matchAndRewrite( |
| 6139 | cir::MatrixExtractOp op, OpAdaptor adaptor, |
| 6140 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6141 | mlir::Value flatIndex = |
| 6142 | computeMatrixFlatIndex(rewriter, loc: op.getLoc(), matrixTy: op.getMatrix().getType(), |
| 6143 | rowIdx: adaptor.getRowIdx(), columnIdx: adaptor.getColumnIdx()); |
| 6144 | rewriter.replaceOpWithNewOp<mlir::LLVM::ExtractElementOp>( |
| 6145 | op, args: adaptor.getMatrix(), args&: flatIndex); |
| 6146 | return mlir::success(); |
| 6147 | } |
| 6148 | |
| 6149 | mlir::LogicalResult CIRToLLVMMatrixExtractRowOpLowering::matchAndRewrite( |
| 6150 | cir::MatrixExtractRowOp op, OpAdaptor adaptor, |
| 6151 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6152 | cir::MatrixType matrixTy = op.getMatrix().getType(); |
| 6153 | mlir::Type elementType = |
| 6154 | typeConverter->convertType(t: matrixTy.getElementType()); |
| 6155 | mlir::Type resultTy = |
| 6156 | mlir::VectorType::get(shape: matrixTy.getNumColumns(), elementType); |
| 6157 | mlir::Value rowVec = |
| 6158 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: resultTy); |
| 6159 | |
| 6160 | for (unsigned col = 0; col != matrixTy.getNumColumns(); ++col) { |
| 6161 | mlir::Value columnIdx = mlir::LLVM::ConstantOp::create( |
| 6162 | builder&: rewriter, location: op.getLoc(), res: rewriter.getI32Type(), |
| 6163 | value: rewriter.getIntegerAttr(type: rewriter.getI32Type(), value: col)); |
| 6164 | mlir::Value flatIndex = computeMatrixFlatIndex( |
| 6165 | rewriter, loc: op.getLoc(), matrixTy, rowIdx: adaptor.getRowIdx(), columnIdx); |
| 6166 | |
| 6167 | auto element = mlir::LLVM::ExtractElementOp::create( |
| 6168 | builder&: rewriter, location: op->getLoc(), res: elementType, vector: adaptor.getMatrix(), position: flatIndex); |
| 6169 | rowVec = mlir::LLVM::InsertElementOp::create(builder&: rewriter, location: op->getLoc(), vector: rowVec, |
| 6170 | value: element, position: columnIdx); |
| 6171 | } |
| 6172 | rewriter.replaceOp(op, newValues: rowVec); |
| 6173 | return mlir::success(); |
| 6174 | } |
| 6175 | |
| 6176 | mlir::LogicalResult CIRToLLVMMatrixInsertOpLowering::matchAndRewrite( |
| 6177 | cir::MatrixInsertOp op, OpAdaptor adaptor, |
| 6178 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6179 | mlir::Value flatIndex = |
| 6180 | computeMatrixFlatIndex(rewriter, loc: op.getLoc(), matrixTy: op.getMatrix().getType(), |
| 6181 | rowIdx: adaptor.getRowIdx(), columnIdx: adaptor.getColumnIdx()); |
| 6182 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertElementOp>( |
| 6183 | op, args: adaptor.getMatrix(), args: adaptor.getValue(), args&: flatIndex); |
| 6184 | return mlir::success(); |
| 6185 | } |
| 6186 | |
| 6187 | mlir::LogicalResult CIRToLLVMMatrixColumnMajorLoadOpLowering::matchAndRewrite( |
| 6188 | cir::MatrixColumnMajorLoadOp op, OpAdaptor adaptor, |
| 6189 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6190 | cir::MatrixType resultMatrixTy = op.getResult().getType(); |
| 6191 | mlir::Type resultTy = typeConverter->convertType(t: resultMatrixTy); |
| 6192 | rewriter.replaceOpWithNewOp<mlir::LLVM::MatrixColumnMajorLoadOp>( |
| 6193 | op, args&: resultTy, args: adaptor.getValue(), args: adaptor.getStride(), |
| 6194 | args: rewriter.getBoolAttr(value: op.getIsVolatile()), |
| 6195 | args: rewriter.getI32IntegerAttr(value: resultMatrixTy.getNumRows()), |
| 6196 | args: rewriter.getI32IntegerAttr(value: resultMatrixTy.getNumColumns())); |
| 6197 | return mlir::success(); |
| 6198 | } |
| 6199 | |
| 6200 | mlir::LogicalResult CIRToLLVMMatrixColumnMajorStoreOpLowering::matchAndRewrite( |
| 6201 | cir::MatrixColumnMajorStoreOp op, OpAdaptor adaptor, |
| 6202 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6203 | cir::MatrixType matrixTy = op.getMatrix().getType(); |
| 6204 | rewriter.replaceOpWithNewOp<mlir::LLVM::MatrixColumnMajorStoreOp>( |
| 6205 | op, args: adaptor.getMatrix(), args: adaptor.getValue(), args: adaptor.getStride(), |
| 6206 | args: rewriter.getBoolAttr(value: op.getIsVolatile()), |
| 6207 | args: rewriter.getI32IntegerAttr(value: matrixTy.getNumRows()), |
| 6208 | args: rewriter.getI32IntegerAttr(value: matrixTy.getNumColumns())); |
| 6209 | return mlir::success(); |
| 6210 | } |
| 6211 | |
| 6212 | mlir::LogicalResult CIRToLLVMMatrixMultiplyOpLowering::matchAndRewrite( |
| 6213 | cir::MatrixMultiplyOp op, OpAdaptor adaptor, |
| 6214 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6215 | cir::MatrixType lhsTy = op.getLhs().getType(); |
| 6216 | cir::MatrixType rhsTy = op.getRhs().getType(); |
| 6217 | mlir::Type resultTy = typeConverter->convertType(t: op.getResult().getType()); |
| 6218 | rewriter.replaceOpWithNewOp<mlir::LLVM::MatrixMultiplyOp>( |
| 6219 | op, args&: resultTy, args: adaptor.getLhs(), args: adaptor.getRhs(), args: lhsTy.getNumRows(), |
| 6220 | args: lhsTy.getNumColumns(), args: rhsTy.getNumColumns()); |
| 6221 | return mlir::success(); |
| 6222 | } |
| 6223 | |
| 6224 | mlir::LogicalResult CIRToLLVMMatrixTransposeOpLowering::matchAndRewrite( |
| 6225 | cir::MatrixTransposeOp op, OpAdaptor adaptor, |
| 6226 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6227 | cir::MatrixType resultMatrixTy = op.getValue().getType(); |
| 6228 | mlir::Type resultTy = typeConverter->convertType(t: resultMatrixTy); |
| 6229 | rewriter.replaceOpWithNewOp<mlir::LLVM::MatrixTransposeOp>( |
| 6230 | op, args&: resultTy, args: adaptor.getValue(), args: resultMatrixTy.getNumRows(), |
| 6231 | args: resultMatrixTy.getNumColumns()); |
| 6232 | return mlir::success(); |
| 6233 | } |
| 6234 | |
| 6235 | mlir::LogicalResult CIRToLLVMComplexCreateOpLowering::matchAndRewrite( |
| 6236 | cir::ComplexCreateOp op, OpAdaptor adaptor, |
| 6237 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6238 | mlir::Type complexLLVMTy = |
| 6239 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 6240 | auto initialComplex = |
| 6241 | mlir::LLVM::UndefOp::create(builder&: rewriter, location: op->getLoc(), res: complexLLVMTy); |
| 6242 | |
| 6243 | auto realComplex = mlir::LLVM::InsertValueOp::create( |
| 6244 | builder&: rewriter, location: op->getLoc(), container: initialComplex, value: adaptor.getReal(), |
| 6245 | position: ArrayRef(int64_t{0})); |
| 6246 | |
| 6247 | auto complex = mlir::LLVM::InsertValueOp::create( |
| 6248 | builder&: rewriter, location: op->getLoc(), container: realComplex, value: adaptor.getImag(), |
| 6249 | position: ArrayRef(int64_t{1})); |
| 6250 | |
| 6251 | rewriter.replaceOp(op, newOp: complex); |
| 6252 | return mlir::success(); |
| 6253 | } |
| 6254 | |
| 6255 | mlir::LogicalResult CIRToLLVMComplexRealOpLowering::matchAndRewrite( |
| 6256 | cir::ComplexRealOp op, OpAdaptor adaptor, |
| 6257 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6258 | mlir::Value operand = adaptor.getOperand(); |
| 6259 | if (mlir::isa<cir::ComplexType>(Val: op.getOperand().getType())) { |
| 6260 | operand = rewriter.createOrFold<mlir::LLVM::ExtractValueOp>( |
| 6261 | location: op.getLoc(), args&: operand, args: llvm::ArrayRef<std::int64_t>{0}); |
| 6262 | } |
| 6263 | rewriter.replaceOp(op, newValues: operand); |
| 6264 | return mlir::success(); |
| 6265 | } |
| 6266 | |
| 6267 | mlir::LogicalResult CIRToLLVMComplexImagOpLowering::matchAndRewrite( |
| 6268 | cir::ComplexImagOp op, OpAdaptor adaptor, |
| 6269 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6270 | mlir::Type resultLLVMTy = getTypeConverter()->convertType(t: op.getType()); |
| 6271 | mlir::Value operand = adaptor.getOperand(); |
| 6272 | mlir::Location loc = op.getLoc(); |
| 6273 | |
| 6274 | if (mlir::isa<cir::ComplexType>(Val: op.getOperand().getType())) { |
| 6275 | operand = rewriter.createOrFold<mlir::LLVM::ExtractValueOp>( |
| 6276 | location: loc, args&: operand, args: llvm::ArrayRef<std::int64_t>{1}); |
| 6277 | } else { |
| 6278 | mlir::TypedAttr zeroAttr = rewriter.getZeroAttr(type: resultLLVMTy); |
| 6279 | operand = |
| 6280 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: resultLLVMTy, value: zeroAttr); |
| 6281 | } |
| 6282 | |
| 6283 | rewriter.replaceOp(op, newValues: operand); |
| 6284 | return mlir::success(); |
| 6285 | } |
| 6286 | |
| 6287 | mlir::LogicalResult CIRToLLVMComplexAddOpLowering::matchAndRewrite( |
| 6288 | cir::ComplexAddOp op, OpAdaptor adaptor, |
| 6289 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6290 | mlir::Value lhs = adaptor.getLhs(); |
| 6291 | mlir::Value rhs = adaptor.getRhs(); |
| 6292 | mlir::Location loc = op.getLoc(); |
| 6293 | |
| 6294 | auto complexType = mlir::cast<cir::ComplexType>(Val: op.getLhs().getType()); |
| 6295 | mlir::Type complexElemTy = |
| 6296 | getTypeConverter()->convertType(t: complexType.getElementType()); |
| 6297 | auto lhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6298 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{0})); |
| 6299 | auto lhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6300 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{1})); |
| 6301 | auto rhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6302 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{0})); |
| 6303 | auto rhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6304 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{1})); |
| 6305 | |
| 6306 | mlir::Value newReal; |
| 6307 | mlir::Value newImag; |
| 6308 | if (complexElemTy.isInteger()) { |
| 6309 | newReal = mlir::LLVM::AddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6310 | rhs: rhsReal); |
| 6311 | newImag = mlir::LLVM::AddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6312 | rhs: rhsImag); |
| 6313 | } else { |
| 6314 | assert(!cir::MissingFeatures::fastMathFlags()); |
| 6315 | assert(!cir::MissingFeatures::fpConstraints()); |
| 6316 | newReal = mlir::LLVM::FAddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6317 | rhs: rhsReal); |
| 6318 | newImag = mlir::LLVM::FAddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6319 | rhs: rhsImag); |
| 6320 | } |
| 6321 | |
| 6322 | mlir::Type complexLLVMTy = |
| 6323 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 6324 | auto initialComplex = |
| 6325 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: complexLLVMTy); |
| 6326 | |
| 6327 | auto realComplex = mlir::LLVM::InsertValueOp::create( |
| 6328 | builder&: rewriter, location: op->getLoc(), container: initialComplex, value: newReal, position: ArrayRef(int64_t{0})); |
| 6329 | |
| 6330 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
| 6331 | op, args&: realComplex, args&: newImag, args: ArrayRef(int64_t{1})); |
| 6332 | |
| 6333 | return mlir::success(); |
| 6334 | } |
| 6335 | |
| 6336 | mlir::LogicalResult CIRToLLVMComplexSubOpLowering::matchAndRewrite( |
| 6337 | cir::ComplexSubOp op, OpAdaptor adaptor, |
| 6338 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6339 | mlir::Value lhs = adaptor.getLhs(); |
| 6340 | mlir::Value rhs = adaptor.getRhs(); |
| 6341 | mlir::Location loc = op.getLoc(); |
| 6342 | |
| 6343 | auto complexType = mlir::cast<cir::ComplexType>(Val: op.getLhs().getType()); |
| 6344 | mlir::Type complexElemTy = |
| 6345 | getTypeConverter()->convertType(t: complexType.getElementType()); |
| 6346 | auto lhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6347 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{0})); |
| 6348 | auto lhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6349 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{1})); |
| 6350 | auto rhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6351 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{0})); |
| 6352 | auto rhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6353 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{1})); |
| 6354 | |
| 6355 | mlir::Value newReal; |
| 6356 | mlir::Value newImag; |
| 6357 | if (complexElemTy.isInteger()) { |
| 6358 | newReal = mlir::LLVM::SubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6359 | rhs: rhsReal); |
| 6360 | newImag = mlir::LLVM::SubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6361 | rhs: rhsImag); |
| 6362 | } else { |
| 6363 | assert(!cir::MissingFeatures::fastMathFlags()); |
| 6364 | assert(!cir::MissingFeatures::fpConstraints()); |
| 6365 | newReal = mlir::LLVM::FSubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6366 | rhs: rhsReal); |
| 6367 | newImag = mlir::LLVM::FSubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6368 | rhs: rhsImag); |
| 6369 | } |
| 6370 | |
| 6371 | mlir::Type complexLLVMTy = |
| 6372 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 6373 | auto initialComplex = |
| 6374 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: complexLLVMTy); |
| 6375 | |
| 6376 | auto realComplex = mlir::LLVM::InsertValueOp::create( |
| 6377 | builder&: rewriter, location: op->getLoc(), container: initialComplex, value: newReal, position: ArrayRef(int64_t{0})); |
| 6378 | |
| 6379 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
| 6380 | op, args&: realComplex, args&: newImag, args: ArrayRef(int64_t{1})); |
| 6381 | |
| 6382 | return mlir::success(); |
| 6383 | } |
| 6384 | |
| 6385 | mlir::LogicalResult CIRToLLVMComplexFAddOpLowering::matchAndRewrite( |
| 6386 | cir::ComplexFAddOp op, OpAdaptor adaptor, |
| 6387 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6388 | mlir::Value lhs = adaptor.getLhs(); |
| 6389 | mlir::Value rhs = adaptor.getRhs(); |
| 6390 | mlir::Location loc = op.getLoc(); |
| 6391 | |
| 6392 | auto complexType = mlir::cast<cir::ComplexType>(Val: op.getLhs().getType()); |
| 6393 | mlir::Type complexElemTy = |
| 6394 | getTypeConverter()->convertType(t: complexType.getElementType()); |
| 6395 | auto lhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6396 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{0})); |
| 6397 | auto lhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6398 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{1})); |
| 6399 | auto rhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6400 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{0})); |
| 6401 | auto rhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6402 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{1})); |
| 6403 | |
| 6404 | assert(!cir::MissingFeatures::fastMathFlags()); |
| 6405 | mlir::Value newReal; |
| 6406 | mlir::Value newImag; |
| 6407 | if (cir::FenvAttr fenv = op.getFenvAttr()) { |
| 6408 | newReal = createConstrainedFPIntrinsicCall( |
| 6409 | rewriter, loc, operands: {lhsReal, rhsReal}, llvmResTy: complexElemTy, fenv, |
| 6410 | /*constrainedMnemonic=*/"fadd" , |
| 6411 | /*hasRoundingMode=*/true); |
| 6412 | |
| 6413 | newImag = createConstrainedFPIntrinsicCall( |
| 6414 | rewriter, loc, operands: {lhsImag, rhsImag}, llvmResTy: complexElemTy, fenv, |
| 6415 | /*constrainedMnemonic=*/"fadd" , |
| 6416 | /*hasRoundingMode=*/true); |
| 6417 | } else { |
| 6418 | newReal = mlir::LLVM::FAddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6419 | rhs: rhsReal); |
| 6420 | newImag = mlir::LLVM::FAddOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6421 | rhs: rhsImag); |
| 6422 | } |
| 6423 | |
| 6424 | mlir::Type complexLLVMTy = |
| 6425 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 6426 | auto initialComplex = |
| 6427 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: complexLLVMTy); |
| 6428 | |
| 6429 | auto realComplex = mlir::LLVM::InsertValueOp::create( |
| 6430 | builder&: rewriter, location: op->getLoc(), container: initialComplex, value: newReal, position: ArrayRef(int64_t{0})); |
| 6431 | |
| 6432 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
| 6433 | op, args&: realComplex, args&: newImag, args: ArrayRef(int64_t{1})); |
| 6434 | |
| 6435 | return mlir::success(); |
| 6436 | } |
| 6437 | |
| 6438 | mlir::LogicalResult CIRToLLVMComplexFSubOpLowering::matchAndRewrite( |
| 6439 | cir::ComplexFSubOp op, OpAdaptor adaptor, |
| 6440 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6441 | mlir::Value lhs = adaptor.getLhs(); |
| 6442 | mlir::Value rhs = adaptor.getRhs(); |
| 6443 | mlir::Location loc = op.getLoc(); |
| 6444 | |
| 6445 | auto complexType = mlir::cast<cir::ComplexType>(Val: op.getLhs().getType()); |
| 6446 | mlir::Type complexElemTy = |
| 6447 | getTypeConverter()->convertType(t: complexType.getElementType()); |
| 6448 | auto lhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6449 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{0})); |
| 6450 | auto lhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6451 | builder&: rewriter, location: loc, res: complexElemTy, container: lhs, position: ArrayRef(int64_t{1})); |
| 6452 | auto rhsReal = mlir::LLVM::ExtractValueOp::create( |
| 6453 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{0})); |
| 6454 | auto rhsImag = mlir::LLVM::ExtractValueOp::create( |
| 6455 | builder&: rewriter, location: loc, res: complexElemTy, container: rhs, position: ArrayRef(int64_t{1})); |
| 6456 | |
| 6457 | assert(!cir::MissingFeatures::fastMathFlags()); |
| 6458 | mlir::Value newReal; |
| 6459 | mlir::Value newImag; |
| 6460 | if (cir::FenvAttr fenv = op.getFenvAttr()) { |
| 6461 | newReal = createConstrainedFPIntrinsicCall( |
| 6462 | rewriter, loc, operands: {lhsReal, rhsReal}, llvmResTy: complexElemTy, fenv, |
| 6463 | /*constrainedMnemonic=*/"fsub" , |
| 6464 | /*hasRoundingMode=*/true); |
| 6465 | |
| 6466 | newImag = createConstrainedFPIntrinsicCall( |
| 6467 | rewriter, loc, operands: {lhsImag, rhsImag}, llvmResTy: complexElemTy, fenv, |
| 6468 | /*constrainedMnemonic=*/"fsub" , |
| 6469 | /*hasRoundingMode=*/true); |
| 6470 | } else { |
| 6471 | newReal = mlir::LLVM::FSubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsReal, |
| 6472 | rhs: rhsReal); |
| 6473 | newImag = mlir::LLVM::FSubOp::create(builder&: rewriter, location: loc, res: complexElemTy, lhs: lhsImag, |
| 6474 | rhs: rhsImag); |
| 6475 | } |
| 6476 | |
| 6477 | mlir::Type complexLLVMTy = |
| 6478 | getTypeConverter()->convertType(t: op.getResult().getType()); |
| 6479 | auto initialComplex = |
| 6480 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: complexLLVMTy); |
| 6481 | |
| 6482 | auto realComplex = mlir::LLVM::InsertValueOp::create( |
| 6483 | builder&: rewriter, location: op->getLoc(), container: initialComplex, value: newReal, position: ArrayRef(int64_t{0})); |
| 6484 | |
| 6485 | rewriter.replaceOpWithNewOp<mlir::LLVM::InsertValueOp>( |
| 6486 | op, args&: realComplex, args&: newImag, args: ArrayRef(int64_t{1})); |
| 6487 | |
| 6488 | return mlir::success(); |
| 6489 | } |
| 6490 | |
| 6491 | mlir::IntegerType computeBitfieldIntType(mlir::Type storageType, |
| 6492 | mlir::MLIRContext *context, |
| 6493 | unsigned &storageSize) { |
| 6494 | return TypeSwitch<mlir::Type, mlir::IntegerType>(storageType) |
| 6495 | .Case<cir::ArrayType>(caseFn: [&](cir::ArrayType atTy) { |
| 6496 | storageSize = atTy.getSize() * 8; |
| 6497 | return mlir::IntegerType::get(context, width: storageSize); |
| 6498 | }) |
| 6499 | .Case<cir::IntType>(caseFn: [&](cir::IntType intTy) { |
| 6500 | storageSize = intTy.getWidth(); |
| 6501 | return mlir::IntegerType::get(context, width: storageSize); |
| 6502 | }) |
| 6503 | .Default(defaultFn: [](mlir::Type) -> mlir::IntegerType { |
| 6504 | llvm_unreachable( |
| 6505 | "Either ArrayType or IntType expected for bitfields storage" ); |
| 6506 | }); |
| 6507 | } |
| 6508 | |
| 6509 | mlir::LogicalResult CIRToLLVMSetBitfieldOpLowering::matchAndRewrite( |
| 6510 | cir::SetBitfieldOp op, OpAdaptor adaptor, |
| 6511 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6512 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 6513 | rewriter.setInsertionPoint(op); |
| 6514 | |
| 6515 | cir::BitfieldInfoAttr info = op.getBitfieldInfo(); |
| 6516 | uint64_t size = info.getSize(); |
| 6517 | uint64_t offset = info.getOffset(); |
| 6518 | mlir::Type storageType = info.getStorageType(); |
| 6519 | mlir::MLIRContext *context = storageType.getContext(); |
| 6520 | |
| 6521 | unsigned storageSize = 0; |
| 6522 | |
| 6523 | mlir::IntegerType intType = |
| 6524 | computeBitfieldIntType(storageType, context, storageSize); |
| 6525 | |
| 6526 | mlir::Value srcVal = createIntCast(bld&: rewriter, src: adaptor.getSrc(), dstTy: intType); |
| 6527 | unsigned srcWidth = storageSize; |
| 6528 | mlir::Value resultVal = srcVal; |
| 6529 | |
| 6530 | if (storageSize != size) { |
| 6531 | assert(storageSize > size && "Invalid bitfield size." ); |
| 6532 | |
| 6533 | mlir::Value val = mlir::LLVM::LoadOp::create( |
| 6534 | builder&: rewriter, location: op.getLoc(), type: intType, addr: adaptor.getAddr(), alignment: op.getAlignment(), |
| 6535 | isVolatile: op.getIsVolatile()); |
| 6536 | |
| 6537 | srcVal = |
| 6538 | createAnd(bld&: rewriter, lhs: srcVal, rhs: llvm::APInt::getLowBitsSet(numBits: srcWidth, loBitsSet: size)); |
| 6539 | resultVal = srcVal; |
| 6540 | srcVal = createShL(bld&: rewriter, lhs: srcVal, rhs: offset); |
| 6541 | |
| 6542 | // Mask out the original value. |
| 6543 | val = createAnd(bld&: rewriter, lhs: val, |
| 6544 | rhs: ~llvm::APInt::getBitsSet(numBits: srcWidth, loBit: offset, hiBit: offset + size)); |
| 6545 | |
| 6546 | // Or together the unchanged values and the source value. |
| 6547 | srcVal = mlir::LLVM::OrOp::create(builder&: rewriter, location: op.getLoc(), lhs: val, rhs: srcVal); |
| 6548 | } |
| 6549 | |
| 6550 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: op.getLoc(), value: srcVal, addr: adaptor.getAddr(), |
| 6551 | alignment: op.getAlignment(), isVolatile: op.getIsVolatile()); |
| 6552 | |
| 6553 | mlir::Type resultTy = getTypeConverter()->convertType(t: op.getType()); |
| 6554 | |
| 6555 | if (info.getIsSigned()) { |
| 6556 | assert(size <= storageSize); |
| 6557 | unsigned highBits = storageSize - size; |
| 6558 | |
| 6559 | if (highBits) { |
| 6560 | resultVal = createShL(bld&: rewriter, lhs: resultVal, rhs: highBits); |
| 6561 | resultVal = createAShR(bld&: rewriter, lhs: resultVal, rhs: highBits); |
| 6562 | } |
| 6563 | } |
| 6564 | |
| 6565 | resultVal = createIntCast(bld&: rewriter, src: resultVal, |
| 6566 | dstTy: mlir::cast<mlir::IntegerType>(Val&: resultTy), |
| 6567 | isSigned: info.getIsSigned()); |
| 6568 | |
| 6569 | rewriter.replaceOp(op, newValues: resultVal); |
| 6570 | return mlir::success(); |
| 6571 | } |
| 6572 | |
| 6573 | mlir::LogicalResult CIRToLLVMComplexImagPtrOpLowering::matchAndRewrite( |
| 6574 | cir::ComplexImagPtrOp op, OpAdaptor adaptor, |
| 6575 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6576 | cir::PointerType operandTy = op.getOperand().getType(); |
| 6577 | mlir::Type resultLLVMTy = getTypeConverter()->convertType(t: op.getType()); |
| 6578 | if (!resultLLVMTy) |
| 6579 | return emitTypeConversionError(op, type: op.getType()); |
| 6580 | mlir::Type elementLLVMTy = |
| 6581 | getTypeConverter()->convertType(t: operandTy.getPointee()); |
| 6582 | |
| 6583 | mlir::LLVM::GEPArg gepIndices[2] = {{0}, {1}}; |
| 6584 | mlir::LLVM::GEPNoWrapFlags inboundsNuw = |
| 6585 | mlir::LLVM::GEPNoWrapFlags::inbounds | mlir::LLVM::GEPNoWrapFlags::nuw; |
| 6586 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 6587 | op, args&: resultLLVMTy, args&: elementLLVMTy, args: adaptor.getOperand(), args&: gepIndices, |
| 6588 | args&: inboundsNuw); |
| 6589 | return mlir::success(); |
| 6590 | } |
| 6591 | |
| 6592 | mlir::LogicalResult CIRToLLVMComplexRealPtrOpLowering::matchAndRewrite( |
| 6593 | cir::ComplexRealPtrOp op, OpAdaptor adaptor, |
| 6594 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6595 | cir::PointerType operandTy = op.getOperand().getType(); |
| 6596 | mlir::Type resultLLVMTy = getTypeConverter()->convertType(t: op.getType()); |
| 6597 | if (!resultLLVMTy) |
| 6598 | return emitTypeConversionError(op, type: op.getType()); |
| 6599 | mlir::Type elementLLVMTy = |
| 6600 | getTypeConverter()->convertType(t: operandTy.getPointee()); |
| 6601 | |
| 6602 | mlir::LLVM::GEPArg gepIndices[2] = {0, 0}; |
| 6603 | mlir::LLVM::GEPNoWrapFlags inboundsNuw = |
| 6604 | mlir::LLVM::GEPNoWrapFlags::inbounds | mlir::LLVM::GEPNoWrapFlags::nuw; |
| 6605 | rewriter.replaceOpWithNewOp<mlir::LLVM::GEPOp>( |
| 6606 | op, args&: resultLLVMTy, args&: elementLLVMTy, args: adaptor.getOperand(), args&: gepIndices, |
| 6607 | args&: inboundsNuw); |
| 6608 | return mlir::success(); |
| 6609 | } |
| 6610 | |
| 6611 | mlir::LogicalResult CIRToLLVMGetBitfieldOpLowering::matchAndRewrite( |
| 6612 | cir::GetBitfieldOp op, OpAdaptor adaptor, |
| 6613 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6614 | |
| 6615 | mlir::OpBuilder::InsertionGuard guard(rewriter); |
| 6616 | rewriter.setInsertionPoint(op); |
| 6617 | |
| 6618 | cir::BitfieldInfoAttr info = op.getBitfieldInfo(); |
| 6619 | uint64_t size = info.getSize(); |
| 6620 | uint64_t offset = info.getOffset(); |
| 6621 | mlir::Type storageType = info.getStorageType(); |
| 6622 | mlir::MLIRContext *context = storageType.getContext(); |
| 6623 | unsigned storageSize = 0; |
| 6624 | |
| 6625 | mlir::IntegerType intType = |
| 6626 | computeBitfieldIntType(storageType, context, storageSize); |
| 6627 | |
| 6628 | mlir::Value val = mlir::LLVM::LoadOp::create( |
| 6629 | builder&: rewriter, location: op.getLoc(), type: intType, addr: adaptor.getAddr(), alignment: op.getAlignment(), |
| 6630 | isVolatile: op.getIsVolatile()); |
| 6631 | val = mlir::LLVM::BitcastOp::create(builder&: rewriter, location: op.getLoc(), res: intType, arg: val); |
| 6632 | |
| 6633 | if (info.getIsSigned()) { |
| 6634 | assert(static_cast<unsigned>(offset + size) <= storageSize); |
| 6635 | unsigned highBits = storageSize - offset - size; |
| 6636 | val = createShL(bld&: rewriter, lhs: val, rhs: highBits); |
| 6637 | val = createAShR(bld&: rewriter, lhs: val, rhs: offset + highBits); |
| 6638 | } else { |
| 6639 | val = createLShR(bld&: rewriter, lhs: val, rhs: offset); |
| 6640 | |
| 6641 | if (static_cast<unsigned>(offset) + size < storageSize) |
| 6642 | val = createAnd(bld&: rewriter, lhs: val, |
| 6643 | rhs: llvm::APInt::getLowBitsSet(numBits: storageSize, loBitsSet: size)); |
| 6644 | } |
| 6645 | |
| 6646 | mlir::Type resTy = getTypeConverter()->convertType(t: op.getType()); |
| 6647 | mlir::Value newOp = createIntCast( |
| 6648 | bld&: rewriter, src: val, dstTy: mlir::cast<mlir::IntegerType>(Val&: resTy), isSigned: info.getIsSigned()); |
| 6649 | rewriter.replaceOp(op, newValues: newOp); |
| 6650 | return mlir::success(); |
| 6651 | } |
| 6652 | |
| 6653 | mlir::LogicalResult CIRToLLVMInlineAsmOpLowering::matchAndRewrite( |
| 6654 | cir::InlineAsmOp op, OpAdaptor adaptor, |
| 6655 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6656 | mlir::Type llResTy; |
| 6657 | if (op.getNumResults()) { |
| 6658 | llResTy = getTypeConverter()->convertType(t: op.getType(i: 0)); |
| 6659 | if (!llResTy) |
| 6660 | return emitTypeConversionError(op, type: op.getType(i: 0)); |
| 6661 | } |
| 6662 | |
| 6663 | cir::AsmFlavor dialect = op.getAsmFlavor(); |
| 6664 | mlir::LLVM::AsmDialect llDialect = dialect == cir::AsmFlavor::x86_att |
| 6665 | ? mlir::LLVM::AsmDialect::AD_ATT |
| 6666 | : mlir::LLVM::AsmDialect::AD_Intel; |
| 6667 | |
| 6668 | SmallVector<mlir::Attribute> opAttrs; |
| 6669 | StringRef llvmAttrName = mlir::LLVM::InlineAsmOp::getElementTypeAttrName(); |
| 6670 | |
| 6671 | // this is for the lowering to LLVM from LLVM dialect. Otherwise, if we |
| 6672 | // don't have the result (i.e. void type as a result of operation), the |
| 6673 | // element type attribute will be attached to the whole instruction, but not |
| 6674 | // to the operand |
| 6675 | if (!op.getNumResults()) |
| 6676 | opAttrs.push_back(Elt: mlir::Attribute()); |
| 6677 | |
| 6678 | // so far we infer the llvm dialect element type attr from |
| 6679 | // CIR operand type. |
| 6680 | for (auto [index, cirOp] : llvm::enumerate(First: op->getOperands())) { |
| 6681 | if (!op.isMemoryOperand(index)) { |
| 6682 | opAttrs.push_back(Elt: mlir::Attribute()); |
| 6683 | continue; |
| 6684 | } |
| 6685 | |
| 6686 | llvm::SmallVector<mlir::NamedAttribute, 1> attrs; |
| 6687 | cir::PointerType typ = mlir::cast<cir::PointerType>(Val: cirOp.getType()); |
| 6688 | mlir::Type elementTy = |
| 6689 | convertTypeForMemory(converter: *getTypeConverter(), dataLayout, type: typ.getPointee()); |
| 6690 | if (!elementTy) |
| 6691 | return emitTypeConversionError(op, type: typ.getPointee()); |
| 6692 | mlir::TypeAttr typAttr = mlir::TypeAttr::get(type: elementTy); |
| 6693 | |
| 6694 | attrs.push_back(Elt: rewriter.getNamedAttr(name: llvmAttrName, val: typAttr)); |
| 6695 | mlir::DictionaryAttr newDict = rewriter.getDictionaryAttr(value: attrs); |
| 6696 | opAttrs.push_back(Elt: newDict); |
| 6697 | } |
| 6698 | |
| 6699 | rewriter.replaceOpWithNewOp<mlir::LLVM::InlineAsmOp>( |
| 6700 | op, args&: llResTy, args: adaptor.getOperands(), args: op.getAsmStringAttr(), |
| 6701 | args: op.getConstraintsAttr(), args: op.getSideEffectsAttr(), |
| 6702 | /*is_align_stack*/ args: mlir::UnitAttr(), |
| 6703 | /*tail_call_kind*/ |
| 6704 | args: mlir::LLVM::TailCallKindAttr::get( |
| 6705 | context: getContext(), tailCallKind: mlir::LLVM::tailcallkind::TailCallKind::None), |
| 6706 | args: op.getConvergentAttr(), |
| 6707 | args: mlir::LLVM::AsmDialectAttr::get(context: getContext(), val: llDialect), |
| 6708 | args: rewriter.getArrayAttr(value: opAttrs)); |
| 6709 | |
| 6710 | return mlir::success(); |
| 6711 | } |
| 6712 | |
| 6713 | mlir::LogicalResult CIRToLLVMVAStartOpLowering::matchAndRewrite( |
| 6714 | cir::VAStartOp op, OpAdaptor adaptor, |
| 6715 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6716 | rewriter.replaceOpWithNewOp<mlir::LLVM::VaStartOp>(op, args: adaptor.getArgList()); |
| 6717 | return mlir::success(); |
| 6718 | } |
| 6719 | |
| 6720 | mlir::LogicalResult CIRToLLVMVAEndOpLowering::matchAndRewrite( |
| 6721 | cir::VAEndOp op, OpAdaptor adaptor, |
| 6722 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6723 | rewriter.replaceOpWithNewOp<mlir::LLVM::VaEndOp>(op, args: adaptor.getArgList()); |
| 6724 | return mlir::success(); |
| 6725 | } |
| 6726 | |
| 6727 | mlir::LogicalResult CIRToLLVMVACopyOpLowering::matchAndRewrite( |
| 6728 | cir::VACopyOp op, OpAdaptor adaptor, |
| 6729 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6730 | rewriter.replaceOpWithNewOp<mlir::LLVM::VaCopyOp>(op, args: adaptor.getDstList(), |
| 6731 | args: adaptor.getSrcList()); |
| 6732 | return mlir::success(); |
| 6733 | } |
| 6734 | |
| 6735 | mlir::LogicalResult CIRToLLVMVAArgOpLowering::matchAndRewrite( |
| 6736 | cir::VAArgOp op, OpAdaptor adaptor, |
| 6737 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6738 | assert(!cir::MissingFeatures::vaArgABILowering()); |
| 6739 | mlir::Type llvmType = |
| 6740 | getTypeConverter()->convertType(t: op->getResultTypes().front()); |
| 6741 | if (!llvmType) |
| 6742 | return mlir::failure(); |
| 6743 | |
| 6744 | rewriter.replaceOpWithNewOp<mlir::LLVM::VaArgOp>(op, args&: llvmType, |
| 6745 | args: adaptor.getArgList()); |
| 6746 | return mlir::success(); |
| 6747 | } |
| 6748 | |
| 6749 | mlir::LogicalResult CIRToLLVMLabelOpLowering::matchAndRewrite( |
| 6750 | cir::LabelOp op, OpAdaptor adaptor, |
| 6751 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6752 | mlir::MLIRContext *ctx = rewriter.getContext(); |
| 6753 | mlir::Block *block = op->getBlock(); |
| 6754 | // A BlockTagOp cannot reside in the entry block. The address of the entry |
| 6755 | // block cannot be taken |
| 6756 | if (block->isEntryBlock()) { |
| 6757 | mlir::Block *newBlock = |
| 6758 | rewriter.splitBlock(block: op->getBlock(), before: mlir::Block::iterator(op)); |
| 6759 | rewriter.setInsertionPointToEnd(block); |
| 6760 | mlir::LLVM::BrOp::create(builder&: rewriter, location: op.getLoc(), dest: newBlock); |
| 6761 | } |
| 6762 | auto tagAttr = |
| 6763 | mlir::LLVM::BlockTagAttr::get(context: ctx, id: blockInfoAddr.getTagIndex()); |
| 6764 | rewriter.setInsertionPoint(op); |
| 6765 | |
| 6766 | auto blockTagOp = |
| 6767 | mlir::LLVM::BlockTagOp::create(builder&: rewriter, location: op->getLoc(), tag: tagAttr); |
| 6768 | mlir::LLVM::LLVMFuncOp func = op->getParentOfType<mlir::LLVM::LLVMFuncOp>(); |
| 6769 | auto blockInfoAttr = |
| 6770 | cir::BlockAddrInfoAttr::get(context: ctx, func_name: func.getSymName(), label_name: op.getLabel()); |
| 6771 | blockInfoAddr.mapBlockTag(info: blockInfoAttr, tagOp: blockTagOp); |
| 6772 | rewriter.eraseOp(op); |
| 6773 | |
| 6774 | return mlir::success(); |
| 6775 | } |
| 6776 | |
| 6777 | mlir::LogicalResult CIRToLLVMBlockAddressOpLowering::matchAndRewrite( |
| 6778 | cir::BlockAddressOp op, OpAdaptor adaptor, |
| 6779 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6780 | mlir::MLIRContext *ctx = rewriter.getContext(); |
| 6781 | |
| 6782 | mlir::LLVM::BlockTagOp matchLabel = |
| 6783 | blockInfoAddr.lookupBlockTag(info: op.getBlockAddrInfoAttr()); |
| 6784 | mlir::LLVM::BlockTagAttr tagAttr; |
| 6785 | if (!matchLabel) |
| 6786 | // If the BlockTagOp has not been emitted yet, use a placeholder. |
| 6787 | // This will later be replaced with the correct tag index during |
| 6788 | // `resolveBlockAddressOp`. |
| 6789 | tagAttr = {}; |
| 6790 | else |
| 6791 | tagAttr = matchLabel.getTag(); |
| 6792 | |
| 6793 | auto blkAddr = mlir::LLVM::BlockAddressAttr::get( |
| 6794 | context: rewriter.getContext(), function: op.getBlockAddrInfoAttr().getFunc(), tag: tagAttr); |
| 6795 | rewriter.setInsertionPoint(op); |
| 6796 | auto newOp = mlir::LLVM::BlockAddressOp::create( |
| 6797 | builder&: rewriter, location: op.getLoc(), res: mlir::LLVM::LLVMPointerType::get(context: ctx), block_addr: blkAddr); |
| 6798 | if (!matchLabel) |
| 6799 | blockInfoAddr.addUnresolvedBlockAddress(op: newOp, info: op.getBlockAddrInfoAttr()); |
| 6800 | rewriter.replaceOp(op, newOp); |
| 6801 | return mlir::success(); |
| 6802 | } |
| 6803 | |
| 6804 | mlir::LogicalResult CIRToLLVMIndirectBrOpLowering::matchAndRewrite( |
| 6805 | cir::IndirectBrOp op, OpAdaptor adaptor, |
| 6806 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6807 | |
| 6808 | mlir::Value targetAddr = adaptor.getAddr(); |
| 6809 | |
| 6810 | // If the poison attribute is set, use llvm.mlir.poison as the address. |
| 6811 | // This happens when the block has no predecessors and is essentially |
| 6812 | // unreachable. Do NOT erase the block argument directly, as that violates |
| 6813 | // the MLIR dialect conversion framework contract (the framework tracks block |
| 6814 | // arguments and will clean them up). A block with no predecessors simply |
| 6815 | // produces no PHI node. |
| 6816 | if (op.getPoison()) { |
| 6817 | auto llvmPtrType = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 6818 | targetAddr = |
| 6819 | mlir::LLVM::PoisonOp::create(builder&: rewriter, location: op->getLoc(), res: llvmPtrType); |
| 6820 | } |
| 6821 | |
| 6822 | rewriter.replaceOpWithNewOp<mlir::LLVM::IndirectBrOp>( |
| 6823 | op, args&: targetAddr, args: adaptor.getSuccOperands(), args: op.getSuccessors()); |
| 6824 | return mlir::success(); |
| 6825 | } |
| 6826 | |
| 6827 | mlir::LogicalResult CIRToLLVMTokenNoneOpLowering::matchAndRewrite( |
| 6828 | cir::TokenNoneOp op, OpAdaptor adaptor, |
| 6829 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6830 | rewriter.replaceOpWithNewOp<mlir::LLVM::NoneTokenOp>( |
| 6831 | op, args: mlir::TokenType::get(ctx: rewriter.getContext())); |
| 6832 | return mlir::success(); |
| 6833 | } |
| 6834 | |
| 6835 | mlir::LogicalResult CIRToLLVMCpuIdOpLowering::matchAndRewrite( |
| 6836 | cir::CpuIdOp op, OpAdaptor adaptor, |
| 6837 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6838 | mlir::ModuleOp moduleOp = op->getParentOfType<mlir::ModuleOp>(); |
| 6839 | std::optional<llvm::Triple> triple = cir::getTripleFromModule(mod: moduleOp); |
| 6840 | if (!triple) |
| 6841 | return op.emitOpError() |
| 6842 | << "requires a '" << cir::CIRDialect::getTripleAttrName() |
| 6843 | << "' module attribute" ; |
| 6844 | |
| 6845 | mlir::Type i32Ty = rewriter.getI32Type(); |
| 6846 | mlir::Type i64Ty = rewriter.getI64Type(); |
| 6847 | mlir::Type i32PtrTy = mlir::LLVM::LLVMPointerType::get(context: i32Ty.getContext(), addressSpace: 0); |
| 6848 | |
| 6849 | mlir::Type cpuidRetTy = mlir::LLVM::LLVMStructType::getLiteral( |
| 6850 | context: rewriter.getContext(), types: {i32Ty, i32Ty, i32Ty, i32Ty}); |
| 6851 | |
| 6852 | mlir::Value functionId = adaptor.getFunctionId(); |
| 6853 | mlir::Value subFunctionId = adaptor.getSubFunctionId(); |
| 6854 | |
| 6855 | StringRef asmString, constraints; |
| 6856 | if (triple->getArch() == llvm::Triple::x86) { |
| 6857 | asmString = "cpuid" ; |
| 6858 | constraints = "={ax},={bx},={cx},={dx},{ax},{cx}" ; |
| 6859 | } else { |
| 6860 | // x86-64 uses %rbx as the base register, so preserve it. |
| 6861 | asmString = "xchgq %rbx, ${1:q}\n" |
| 6862 | "cpuid\n" |
| 6863 | "xchgq %rbx, ${1:q}" ; |
| 6864 | constraints = "={ax},=r,={cx},={dx},0,2" ; |
| 6865 | } |
| 6866 | |
| 6867 | mlir::Value inlineAsm = |
| 6868 | mlir::LLVM::InlineAsmOp::create( |
| 6869 | builder&: rewriter, location: op.getLoc(), res: cpuidRetTy, operands: {functionId, subFunctionId}, |
| 6870 | asm_string: rewriter.getStringAttr(bytes: asmString), |
| 6871 | constraints: rewriter.getStringAttr(bytes: constraints), |
| 6872 | /*has_side_effects=*/mlir::UnitAttr{}, |
| 6873 | /*is_align_stack=*/mlir::UnitAttr{}, |
| 6874 | /*tail_call_kind=*/mlir::LLVM::TailCallKindAttr{}, |
| 6875 | /*convergent=*/{}, |
| 6876 | /*asm_dialect=*/mlir::LLVM::AsmDialectAttr{}, |
| 6877 | /*operand_attrs=*/mlir::ArrayAttr{}) |
| 6878 | .getResult(i: 0); |
| 6879 | |
| 6880 | mlir::Value basePtr = adaptor.getCpuInfo(); |
| 6881 | |
| 6882 | mlir::DataLayout layout(op->getParentOfType<mlir::ModuleOp>()); |
| 6883 | unsigned alignment = layout.getTypeABIAlignment(t: i32Ty); |
| 6884 | for (unsigned i = 0; i < 4; i++) { |
| 6885 | mlir::Value = |
| 6886 | mlir::LLVM::ExtractValueOp::create(builder&: rewriter, location: op.getLoc(), container: inlineAsm, position: i) |
| 6887 | .getResult(); |
| 6888 | mlir::Value index = mlir::LLVM::ConstantOp::create( |
| 6889 | builder&: rewriter, location: op.getLoc(), res: i64Ty, value: rewriter.getI64IntegerAttr(value: i)); |
| 6890 | llvm::SmallVector<mlir::Value, 1> gepIndices = {index}; |
| 6891 | mlir::Value storePtr = mlir::LLVM::GEPOp::create( |
| 6892 | builder&: rewriter, location: op.getLoc(), resultType: i32PtrTy, elementType: i32Ty, basePtr, |
| 6893 | indices: gepIndices, noWrapFlags: mlir::LLVM::GEPNoWrapFlags::none) |
| 6894 | .getResult(); |
| 6895 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: op.getLoc(), value: extracted, addr: storePtr, |
| 6896 | alignment); |
| 6897 | } |
| 6898 | |
| 6899 | rewriter.eraseOp(op); |
| 6900 | return mlir::success(); |
| 6901 | } |
| 6902 | |
| 6903 | mlir::LogicalResult CIRToLLVMMemChrOpLowering::matchAndRewrite( |
| 6904 | cir::MemChrOp op, OpAdaptor adaptor, |
| 6905 | mlir::ConversionPatternRewriter &rewriter) const { |
| 6906 | auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 6907 | mlir::Type srcTy = getTypeConverter()->convertType(t: op.getSrc().getType()); |
| 6908 | mlir::Type patternTy = |
| 6909 | getTypeConverter()->convertType(t: op.getPattern().getType()); |
| 6910 | mlir::Type lenTy = getTypeConverter()->convertType(t: op.getLen().getType()); |
| 6911 | auto fnTy = |
| 6912 | mlir::LLVM::LLVMFunctionType::get(result: llvmPtrTy, arguments: {srcTy, patternTy, lenTy}, |
| 6913 | /*isVarArg=*/false); |
| 6914 | llvm::StringRef fnName = "memchr" ; |
| 6915 | |
| 6916 | mlir::Builder b(rewriter.getContext()); |
| 6917 | mlir::NamedAttribute noundefAttr = |
| 6918 | b.getNamedAttr(name: "llvm.noundef" , val: b.getUnitAttr()); |
| 6919 | mlir::DictionaryAttr noundefDict = mlir::DictionaryAttr::get( |
| 6920 | context: rewriter.getContext(), value: llvm::ArrayRef(noundefAttr)); |
| 6921 | SmallVector<mlir::Attribute> argAttrVec(3, noundefDict); |
| 6922 | mlir::ArrayAttr argAttrs = |
| 6923 | mlir::ArrayAttr::get(context: rewriter.getContext(), value: argAttrVec); |
| 6924 | |
| 6925 | createLLVMFuncOpIfNotExist(rewriter, symbolTables, srcOp: op, fnName, fnTy, |
| 6926 | argAttrs); |
| 6927 | |
| 6928 | mlir::LLVM::CallOp newCall = rewriter.replaceOpWithNewOp<mlir::LLVM::CallOp>( |
| 6929 | op, args: mlir::TypeRange{llvmPtrTy}, args&: fnName, |
| 6930 | args: mlir::ValueRange{adaptor.getSrc(), adaptor.getPattern(), |
| 6931 | adaptor.getLen()}); |
| 6932 | newCall.setArgAttrsAttr(argAttrs); |
| 6933 | return mlir::success(); |
| 6934 | } |
| 6935 | |
| 6936 | // Function to do the clear-padding operation. This is a faithful translation of |
| 6937 | // CGBuiltin.cpp's ClearPadding function. |
| 6938 | static void clearPadding(mlir::ConversionPatternRewriter &rewriter, |
| 6939 | mlir::Location loc, mlir::Value inputPtr, |
| 6940 | uint64_t baseAlignment, |
| 6941 | cir::OffsetPairAttr paddingAttr) { |
| 6942 | // FIXME(cir): Classic-codegen pulls this from the data layout, but MLIR |
| 6943 | // DataLayout just assumes it is 8 everywhere (as does all our lowering). |
| 6944 | // Clang doesn't support any non-8-bit-CHAR_BIT architectures, but keeping |
| 6945 | // this separate so we can use it next time. |
| 6946 | uint64_t charWidth = 8; |
| 6947 | mlir::Type i8Ty = rewriter.getI8Type(); |
| 6948 | mlir::Type ptrTy = mlir::LLVM::LLVMPointerType::get(context: i8Ty.getContext()); |
| 6949 | |
| 6950 | auto startByte = paddingAttr.getStart() / charWidth; |
| 6951 | auto startBit = paddingAttr.getStart() % charWidth; |
| 6952 | auto endByte = paddingAttr.getEnd() / charWidth; |
| 6953 | auto endBit = paddingAttr.getEnd() % charWidth; |
| 6954 | |
| 6955 | if (startByte == endByte) { |
| 6956 | // Interval is within a single byte |
| 6957 | auto index = mlir::LLVM::ConstantOp::create( |
| 6958 | builder&: rewriter, location: loc, type: rewriter.getI32Type(), value: startByte); |
| 6959 | auto element = mlir::LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: ptrTy, elementType: i8Ty, |
| 6960 | basePtr: inputPtr, indices: {index}); |
| 6961 | |
| 6962 | uint64_t adjustedAlignment = llvm::MinAlign(A: baseAlignment, B: startByte); |
| 6963 | |
| 6964 | auto value = mlir::LLVM::LoadOp::create(builder&: rewriter, location: loc, type: i8Ty, addr: element, |
| 6965 | alignment: adjustedAlignment); |
| 6966 | |
| 6967 | // Create mask to clear bits within the byte |
| 6968 | // We want to clear bits from StartBit to EndBit-1 |
| 6969 | uint8_t bitsToClear = ((1 << endBit) - 1) & ~((1 << startBit) - 1); |
| 6970 | uint8_t bitsToKeep = ~bitsToClear; |
| 6971 | auto maskValue = |
| 6972 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: i8Ty, value: bitsToKeep); |
| 6973 | auto newValue = mlir::LLVM::AndOp::create(builder&: rewriter, location: loc, lhs: value, rhs: maskValue); |
| 6974 | |
| 6975 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: loc, value: newValue, addr: element, |
| 6976 | alignment: adjustedAlignment); |
| 6977 | } else { |
| 6978 | // Handle the start byte |
| 6979 | if (startBit != 0) { |
| 6980 | auto index = mlir::LLVM::ConstantOp::create( |
| 6981 | builder&: rewriter, location: loc, type: rewriter.getI32Type(), value: startByte); |
| 6982 | auto element = mlir::LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: ptrTy, elementType: i8Ty, |
| 6983 | basePtr: inputPtr, indices: {index}); |
| 6984 | uint64_t adjustedAlignment = llvm::MinAlign(A: baseAlignment, B: startByte); |
| 6985 | |
| 6986 | auto value = mlir::LLVM::LoadOp::create(builder&: rewriter, location: loc, type: i8Ty, addr: element, |
| 6987 | alignment: adjustedAlignment); |
| 6988 | |
| 6989 | uint8_t bitsToClear = ((1 << (charWidth - startBit)) - 1) << startBit; |
| 6990 | uint8_t bitsToKeep = ~bitsToClear; |
| 6991 | auto maskValue = |
| 6992 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: i8Ty, value: bitsToKeep); |
| 6993 | auto newValue = |
| 6994 | mlir::LLVM::AndOp::create(builder&: rewriter, location: loc, lhs: value, rhs: maskValue); |
| 6995 | |
| 6996 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: loc, value: newValue, addr: element, |
| 6997 | alignment: adjustedAlignment); |
| 6998 | ++startByte; |
| 6999 | } |
| 7000 | |
| 7001 | // Handle full bytes in the middle |
| 7002 | for (auto offset = startByte; offset < endByte; ++offset) { |
| 7003 | auto index = mlir::LLVM::ConstantOp::create( |
| 7004 | builder&: rewriter, location: loc, type: rewriter.getI32Type(), value: offset); |
| 7005 | auto element = mlir::LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: ptrTy, elementType: i8Ty, |
| 7006 | basePtr: inputPtr, indices: {index}); |
| 7007 | uint64_t adjustedAlignment = llvm::MinAlign(A: baseAlignment, B: offset); |
| 7008 | |
| 7009 | auto zero = mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: i8Ty, value: 0); |
| 7010 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: loc, value: zero, addr: element, |
| 7011 | alignment: adjustedAlignment); |
| 7012 | } |
| 7013 | |
| 7014 | // Handle the end byte |
| 7015 | if (endBit != 0) { |
| 7016 | auto index = mlir::LLVM::ConstantOp::create( |
| 7017 | builder&: rewriter, location: loc, type: rewriter.getI32Type(), value: endByte); |
| 7018 | auto element = mlir::LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: ptrTy, elementType: i8Ty, |
| 7019 | basePtr: inputPtr, indices: {index}); |
| 7020 | uint64_t adjustedAlignment = llvm::MinAlign(A: baseAlignment, B: endByte); |
| 7021 | |
| 7022 | auto value = mlir::LLVM::LoadOp::create(builder&: rewriter, location: loc, type: i8Ty, addr: element, |
| 7023 | alignment: adjustedAlignment); |
| 7024 | |
| 7025 | uint8_t bitsToClear = (1 << endBit) - 1; |
| 7026 | uint8_t bitsToKeep = ~bitsToClear; |
| 7027 | auto maskValue = |
| 7028 | mlir::LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: i8Ty, value: bitsToKeep); |
| 7029 | auto newValue = |
| 7030 | mlir::LLVM::AndOp::create(builder&: rewriter, location: loc, lhs: value, rhs: maskValue); |
| 7031 | |
| 7032 | mlir::LLVM::StoreOp::create(builder&: rewriter, location: loc, value: newValue, addr: element, |
| 7033 | alignment: adjustedAlignment); |
| 7034 | } |
| 7035 | } |
| 7036 | } |
| 7037 | |
| 7038 | mlir::LogicalResult CIRToLLVMClearPaddingOpLowering::matchAndRewrite( |
| 7039 | cir::ClearPaddingOp op, OpAdaptor adaptor, |
| 7040 | mlir::ConversionPatternRewriter &rewriter) const { |
| 7041 | |
| 7042 | mlir::Value inputPtr = adaptor.getArg(); |
| 7043 | for (mlir::Attribute attr : op.getPadding()) |
| 7044 | clearPadding(rewriter, loc: op.getLoc(), inputPtr, baseAlignment: op.getAlignment(), |
| 7045 | paddingAttr: cast<cir::OffsetPairAttr>(Val&: attr)); |
| 7046 | |
| 7047 | rewriter.eraseOp(op); |
| 7048 | return mlir::success(); |
| 7049 | } |
| 7050 | |
| 7051 | std::unique_ptr<mlir::Pass> createConvertCIRToLLVMPass() { |
| 7052 | return std::make_unique<ConvertCIRToLLVMPass>(); |
| 7053 | } |
| 7054 | |
| 7055 | void populateCIRToLLVMPasses(mlir::OpPassManager &pm, bool enableOpenMP, |
| 7056 | bool enableOptimizations) { |
| 7057 | mlir::populateCIRPreLoweringPasses(pm); |
| 7058 | if (enableOpenMP) |
| 7059 | pm.addPass(pass: mlir::omp::createMarkDeclareTargetPass()); |
| 7060 | pm.addPass(pass: createConvertCIRToLLVMPass()); |
| 7061 | if (enableOpenMP) { |
| 7062 | pm.addPass(pass: mlir::omp::createHostOpFilteringPass()); |
| 7063 | pm.nest<mlir::LLVM::LLVMFuncOp>().addPass( |
| 7064 | pass: mlir::omp::createStackToSharedPass()); |
| 7065 | } |
| 7066 | if (enableOptimizations) |
| 7067 | pm.nest<mlir::LLVM::LLVMFuncOp>().addPass( |
| 7068 | pass: mlir::LLVM::createLLVMSimplifyCFGPass()); |
| 7069 | } |
| 7070 | |
| 7071 | std::unique_ptr<llvm::Module> |
| 7072 | lowerDirectlyFromCIRToLLVMIR(mlir::ModuleOp mlirModule, LLVMContext &llvmCtx, |
| 7073 | bool enableOpenMP, StringRef mlirSaveTempsOutFile, |
| 7074 | llvm::vfs::FileSystem *fs, |
| 7075 | bool enableOptimizations) { |
| 7076 | llvm::TimeTraceScope scope("lower from CIR to LLVM directly" ); |
| 7077 | |
| 7078 | mlir::MLIRContext *mlirCtx = mlirModule.getContext(); |
| 7079 | |
| 7080 | mlir::PassManager pm(mlirCtx); |
| 7081 | populateCIRToLLVMPasses(pm, enableOpenMP, enableOptimizations); |
| 7082 | |
| 7083 | (void)mlir::applyPassManagerCLOptions(pm); |
| 7084 | |
| 7085 | // The failing pass has already reported its error through the MLIR |
| 7086 | // diagnostic engine. |
| 7087 | if (mlir::failed(Result: pm.run(op: mlirModule))) |
| 7088 | return nullptr; |
| 7089 | |
| 7090 | if (!mlirSaveTempsOutFile.empty()) { |
| 7091 | std::error_code ec; |
| 7092 | llvm::raw_fd_ostream out(mlirSaveTempsOutFile, ec); |
| 7093 | if (!ec) |
| 7094 | mlirModule->print(os&: out); |
| 7095 | } |
| 7096 | |
| 7097 | mlir::registerBuiltinDialectTranslation(context&: *mlirCtx); |
| 7098 | mlir::registerLLVMDialectTranslation(context&: *mlirCtx); |
| 7099 | mlir::registerOpenMPDialectTranslation(context&: *mlirCtx); |
| 7100 | mlir::registerCIRDialectTranslation(context&: *mlirCtx); |
| 7101 | |
| 7102 | llvm::TimeTraceScope translateScope("translateModuleToLLVMIR" ); |
| 7103 | |
| 7104 | StringRef moduleName = mlirModule.getName().value_or(u: "CIRToLLVMModule" ); |
| 7105 | std::unique_ptr<llvm::Module> llvmModule = mlir::translateModuleToLLVMIR( |
| 7106 | module: mlirModule, llvmContext&: llvmCtx, name: moduleName, /*disableVerification=*/false, fs); |
| 7107 | |
| 7108 | return llvmModule; |
| 7109 | } |
| 7110 | } // namespace direct |
| 7111 | } // namespace cir |
| 7112 | |