| 1 | //===- MemRefToLLVM.cpp - MemRef to LLVM dialect conversion ---------------===// |
| 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 | #include "mlir/Conversion/MemRefToLLVM/MemRefToLLVM.h" |
| 10 | |
| 11 | #include "mlir/Analysis/DataLayoutAnalysis.h" |
| 12 | #include "mlir/Conversion/ConvertToLLVM/ToLLVMInterface.h" |
| 13 | #include "mlir/Conversion/LLVMCommon/ConversionTarget.h" |
| 14 | #include "mlir/Conversion/LLVMCommon/Pattern.h" |
| 15 | #include "mlir/Conversion/LLVMCommon/TypeConverter.h" |
| 16 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 17 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 18 | #include "mlir/Dialect/LLVMIR/FunctionCallUtils.h" |
| 19 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" |
| 20 | #include "mlir/Dialect/LLVMIR/LLVMTypes.h" |
| 21 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 22 | #include "mlir/Dialect/MemRef/Utils/MemRefUtils.h" |
| 23 | #include "mlir/IR/AffineMap.h" |
| 24 | #include "mlir/IR/BuiltinTypes.h" |
| 25 | #include "mlir/IR/IRMapping.h" |
| 26 | #include "mlir/Pass/Pass.h" |
| 27 | #include "llvm/Support/DebugLog.h" |
| 28 | #include "llvm/Support/MathExtras.h" |
| 29 | |
| 30 | #include <optional> |
| 31 | |
| 32 | #define DEBUG_TYPE "memref-to-llvm" |
| 33 | |
| 34 | namespace mlir { |
| 35 | #define GEN_PASS_DEF_FINALIZEMEMREFTOLLVMCONVERSIONPASS |
| 36 | #include "mlir/Conversion/Passes.h.inc" |
| 37 | } // namespace mlir |
| 38 | |
| 39 | using namespace mlir; |
| 40 | |
| 41 | /// Returns GEP no-wrap flags for a memref load/store. |
| 42 | /// inbounds is always valid when indices are in-bounds per the memref spec. |
| 43 | /// nuw requires every index*stride term to not unsigned-wrap, which holds iff |
| 44 | /// all strides are statically non-negative. Negative strides would make the |
| 45 | /// intermediate mul nuw overflow (e.g., idx * (-1 as u64) wraps for idx > 0). |
| 46 | static LLVM::GEPNoWrapFlags getLoadStoreNoWrapFlags(MemRefType type) { |
| 47 | auto [strides, offset] = type.getStridesAndOffset(); |
| 48 | LLVM::GEPNoWrapFlags flags = LLVM::GEPNoWrapFlags::inbounds; |
| 49 | if (llvm::all_of(Range&: strides, P: [](int64_t s) { |
| 50 | return !ShapedType::isDynamic(dValue: s) && s >= 0; |
| 51 | })) |
| 52 | flags = flags | LLVM::GEPNoWrapFlags::nuw; |
| 53 | return flags; |
| 54 | } |
| 55 | |
| 56 | namespace { |
| 57 | |
| 58 | static bool isStaticStrideOrOffset(int64_t strideOrOffset) { |
| 59 | return ShapedType::isStatic(dValue: strideOrOffset); |
| 60 | } |
| 61 | |
| 62 | static FailureOr<LLVM::LLVMFuncOp> |
| 63 | getFreeFn(OpBuilder &b, const LLVMTypeConverter *typeConverter, |
| 64 | Operation *module, SymbolTableCollection *symbolTables) { |
| 65 | bool useGenericFn = typeConverter->getOptions().useGenericFunctions; |
| 66 | |
| 67 | if (useGenericFn) |
| 68 | return LLVM::lookupOrCreateGenericFreeFn(b, moduleOp: module, symbolTables); |
| 69 | |
| 70 | return LLVM::lookupOrCreateFreeFn(b, moduleOp: module, symbolTables); |
| 71 | } |
| 72 | |
| 73 | static FailureOr<LLVM::LLVMFuncOp> |
| 74 | getNotalignedAllocFn(OpBuilder &b, const LLVMTypeConverter *typeConverter, |
| 75 | Operation *module, Type indexType, |
| 76 | SymbolTableCollection *symbolTables) { |
| 77 | bool useGenericFn = typeConverter->getOptions().useGenericFunctions; |
| 78 | if (useGenericFn) |
| 79 | return LLVM::lookupOrCreateGenericAllocFn(b, moduleOp: module, indexType, |
| 80 | symbolTables); |
| 81 | |
| 82 | return LLVM::lookupOrCreateMallocFn(b, moduleOp: module, indexType, symbolTables); |
| 83 | } |
| 84 | |
| 85 | static FailureOr<LLVM::LLVMFuncOp> |
| 86 | getAlignedAllocFn(OpBuilder &b, const LLVMTypeConverter *typeConverter, |
| 87 | Operation *module, Type indexType, |
| 88 | SymbolTableCollection *symbolTables) { |
| 89 | bool useGenericFn = typeConverter->getOptions().useGenericFunctions; |
| 90 | |
| 91 | if (useGenericFn) |
| 92 | return LLVM::lookupOrCreateGenericAlignedAllocFn(b, moduleOp: module, indexType, |
| 93 | symbolTables); |
| 94 | |
| 95 | return LLVM::lookupOrCreateAlignedAllocFn(b, moduleOp: module, indexType, symbolTables); |
| 96 | } |
| 97 | |
| 98 | /// Computes the aligned value for 'input' as follows: |
| 99 | /// bumped = input + alignement - 1 |
| 100 | /// aligned = bumped - bumped % alignment |
| 101 | static Value createAligned(ConversionPatternRewriter &rewriter, Location loc, |
| 102 | Value input, Value alignment) { |
| 103 | Value one = |
| 104 | LLVM::ConstantOp::create(builder&: rewriter, location: loc, res: alignment.getType(), |
| 105 | value: rewriter.getIntegerAttr(type: alignment.getType(), value: 1)); |
| 106 | Value bump = LLVM::SubOp::create(builder&: rewriter, location: loc, lhs: alignment, rhs: one); |
| 107 | Value bumped = LLVM::AddOp::create(builder&: rewriter, location: loc, lhs: input, rhs: bump); |
| 108 | Value mod = LLVM::URemOp::create(builder&: rewriter, location: loc, lhs: bumped, rhs: alignment); |
| 109 | return LLVM::SubOp::create(builder&: rewriter, location: loc, lhs: bumped, rhs: mod); |
| 110 | } |
| 111 | |
| 112 | /// Computes the byte size for the MemRef element type. |
| 113 | static unsigned getMemRefEltSizeInBytes(const LLVMTypeConverter *typeConverter, |
| 114 | MemRefType memRefType, Operation *op, |
| 115 | const DataLayout *defaultLayout) { |
| 116 | const DataLayout *layout = defaultLayout; |
| 117 | if (const DataLayoutAnalysis *analysis = |
| 118 | typeConverter->getDataLayoutAnalysis()) { |
| 119 | layout = &analysis->getAbove(operation: op); |
| 120 | } |
| 121 | Type elementType = memRefType.getElementType(); |
| 122 | if (auto memRefElementType = dyn_cast<MemRefType>(Val&: elementType)) |
| 123 | return typeConverter->getMemRefDescriptorSize(type: memRefElementType, layout: *layout); |
| 124 | if (auto memRefElementType = dyn_cast<UnrankedMemRefType>(Val&: elementType)) |
| 125 | return typeConverter->getUnrankedMemRefDescriptorSize(type: memRefElementType, |
| 126 | layout: *layout); |
| 127 | return layout->getTypeSize(t: elementType); |
| 128 | } |
| 129 | |
| 130 | static Value castAllocFuncResult(ConversionPatternRewriter &rewriter, |
| 131 | Location loc, Value allocatedPtr, |
| 132 | MemRefType memRefType, Type elementPtrType, |
| 133 | const LLVMTypeConverter &typeConverter) { |
| 134 | auto allocatedPtrTy = cast<LLVM::LLVMPointerType>(Val: allocatedPtr.getType()); |
| 135 | FailureOr<unsigned> maybeMemrefAddrSpace = |
| 136 | typeConverter.getMemRefAddressSpace(type: memRefType); |
| 137 | assert(succeeded(maybeMemrefAddrSpace) && "unsupported address space" ); |
| 138 | unsigned memrefAddrSpace = *maybeMemrefAddrSpace; |
| 139 | if (allocatedPtrTy.getAddressSpace() != memrefAddrSpace) |
| 140 | allocatedPtr = LLVM::AddrSpaceCastOp::create( |
| 141 | builder&: rewriter, location: loc, |
| 142 | res: LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: memrefAddrSpace), |
| 143 | arg: allocatedPtr); |
| 144 | return allocatedPtr; |
| 145 | } |
| 146 | |
| 147 | class AllocOpLowering : public ConvertOpToLLVMPattern<memref::AllocOp> { |
| 148 | SymbolTableCollection *symbolTables = nullptr; |
| 149 | |
| 150 | public: |
| 151 | explicit AllocOpLowering(const LLVMTypeConverter &typeConverter, |
| 152 | SymbolTableCollection *symbolTables = nullptr, |
| 153 | PatternBenefit benefit = 1) |
| 154 | : ConvertOpToLLVMPattern<memref::AllocOp>(typeConverter, benefit), |
| 155 | symbolTables(symbolTables) {} |
| 156 | |
| 157 | LogicalResult |
| 158 | matchAndRewrite(memref::AllocOp op, OpAdaptor adaptor, |
| 159 | ConversionPatternRewriter &rewriter) const override { |
| 160 | auto loc = op.getLoc(); |
| 161 | MemRefType memRefType = op.getType(); |
| 162 | if (!isConvertibleAndHasIdentityMaps(type: memRefType)) |
| 163 | return rewriter.notifyMatchFailure(arg&: op, msg: "incompatible memref type" ); |
| 164 | |
| 165 | // Get or insert alloc function into the module. |
| 166 | FailureOr<LLVM::LLVMFuncOp> allocFuncOp = |
| 167 | getNotalignedAllocFn(b&: rewriter, typeConverter: getTypeConverter(), |
| 168 | module: op->getParentWithTrait<OpTrait::SymbolTable>(), |
| 169 | indexType: getIndexType(), symbolTables); |
| 170 | if (failed(Result: allocFuncOp)) |
| 171 | return failure(); |
| 172 | |
| 173 | // Get actual sizes of the memref as values: static sizes are constant |
| 174 | // values and dynamic sizes are passed to 'alloc' as operands. In case of |
| 175 | // zero-dimensional memref, assume a scalar (size 1). |
| 176 | SmallVector<Value, 4> sizes; |
| 177 | SmallVector<Value, 4> strides; |
| 178 | Value sizeBytes; |
| 179 | |
| 180 | this->getMemRefDescriptorSizes(loc, memRefType, dynamicSizes: adaptor.getOperands(), |
| 181 | rewriter, sizes, strides, size&: sizeBytes, sizeInBytes: true); |
| 182 | |
| 183 | Value alignment = getAlignment(rewriter, loc, op); |
| 184 | if (alignment) { |
| 185 | // Adjust the allocation size to consider alignment. |
| 186 | sizeBytes = LLVM::AddOp::create(builder&: rewriter, location: loc, lhs: sizeBytes, rhs: alignment); |
| 187 | } |
| 188 | |
| 189 | // Allocate the underlying buffer. |
| 190 | Type elementPtrType = this->getElementPtrType(type: memRefType); |
| 191 | assert(elementPtrType && "could not compute element ptr type" ); |
| 192 | auto results = |
| 193 | LLVM::CallOp::create(builder&: rewriter, location: loc, func: allocFuncOp.value(), args: sizeBytes); |
| 194 | |
| 195 | Value allocatedPtr = |
| 196 | castAllocFuncResult(rewriter, loc, allocatedPtr: results.getResult(), memRefType, |
| 197 | elementPtrType, typeConverter: *getTypeConverter()); |
| 198 | Value alignedPtr = allocatedPtr; |
| 199 | if (alignment) { |
| 200 | // Compute the aligned pointer. |
| 201 | Value allocatedInt = |
| 202 | LLVM::PtrToIntOp::create(builder&: rewriter, location: loc, res: getIndexType(), arg: allocatedPtr); |
| 203 | Value alignmentInt = |
| 204 | createAligned(rewriter, loc, input: allocatedInt, alignment); |
| 205 | alignedPtr = |
| 206 | LLVM::IntToPtrOp::create(builder&: rewriter, location: loc, resultType: elementPtrType, operands: alignmentInt); |
| 207 | } |
| 208 | |
| 209 | // Create the MemRef descriptor. |
| 210 | auto memRefDescriptor = this->createMemRefDescriptor( |
| 211 | loc, memRefType, allocatedPtr, alignedPtr, sizes, strides, rewriter); |
| 212 | |
| 213 | // Return the final value of the descriptor. |
| 214 | rewriter.replaceOp(op, newValues: {memRefDescriptor}); |
| 215 | return success(); |
| 216 | } |
| 217 | |
| 218 | /// Computes the alignment for the given memory allocation op. |
| 219 | template <typename OpType> |
| 220 | Value getAlignment(ConversionPatternRewriter &rewriter, Location loc, |
| 221 | OpType op) const { |
| 222 | MemRefType memRefType = op.getType(); |
| 223 | Value alignment; |
| 224 | if (auto alignmentAttr = op.getAlignment()) { |
| 225 | Type indexType = getIndexType(); |
| 226 | alignment = |
| 227 | createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: *alignmentAttr); |
| 228 | } else if (!memRefType.getElementType().isSignlessIntOrIndexOrFloat()) { |
| 229 | // In the case where no alignment is specified, we may want to override |
| 230 | // `malloc's` behavior. `malloc` typically aligns at the size of the |
| 231 | // biggest scalar on a target HW. For non-scalars, use the natural |
| 232 | // alignment of the LLVM type given by the LLVM DataLayout. |
| 233 | alignment = getSizeInBytes(loc, type: memRefType.getElementType(), rewriter); |
| 234 | } |
| 235 | return alignment; |
| 236 | } |
| 237 | }; |
| 238 | |
| 239 | class AlignedAllocOpLowering : public ConvertOpToLLVMPattern<memref::AllocOp> { |
| 240 | SymbolTableCollection *symbolTables = nullptr; |
| 241 | |
| 242 | public: |
| 243 | explicit AlignedAllocOpLowering(const LLVMTypeConverter &typeConverter, |
| 244 | SymbolTableCollection *symbolTables = nullptr, |
| 245 | PatternBenefit benefit = 1) |
| 246 | : ConvertOpToLLVMPattern<memref::AllocOp>(typeConverter, benefit), |
| 247 | symbolTables(symbolTables) {} |
| 248 | |
| 249 | LogicalResult |
| 250 | matchAndRewrite(memref::AllocOp op, OpAdaptor adaptor, |
| 251 | ConversionPatternRewriter &rewriter) const override { |
| 252 | auto loc = op.getLoc(); |
| 253 | MemRefType memRefType = op.getType(); |
| 254 | if (!isConvertibleAndHasIdentityMaps(type: memRefType)) |
| 255 | return rewriter.notifyMatchFailure(arg&: op, msg: "incompatible memref type" ); |
| 256 | |
| 257 | // Get or insert alloc function into module. |
| 258 | FailureOr<LLVM::LLVMFuncOp> allocFuncOp = |
| 259 | getAlignedAllocFn(b&: rewriter, typeConverter: getTypeConverter(), |
| 260 | module: op->getParentWithTrait<OpTrait::SymbolTable>(), |
| 261 | indexType: getIndexType(), symbolTables); |
| 262 | if (failed(Result: allocFuncOp)) |
| 263 | return failure(); |
| 264 | |
| 265 | // Get actual sizes of the memref as values: static sizes are constant |
| 266 | // values and dynamic sizes are passed to 'alloc' as operands. In case of |
| 267 | // zero-dimensional memref, assume a scalar (size 1). |
| 268 | SmallVector<Value, 4> sizes; |
| 269 | SmallVector<Value, 4> strides; |
| 270 | Value sizeBytes; |
| 271 | |
| 272 | this->getMemRefDescriptorSizes(loc, memRefType, dynamicSizes: adaptor.getOperands(), |
| 273 | rewriter, sizes, strides, size&: sizeBytes, sizeInBytes: !false); |
| 274 | |
| 275 | int64_t alignment = alignedAllocationGetAlignment(op, defaultLayout: &defaultLayout); |
| 276 | |
| 277 | Value allocAlignment = |
| 278 | createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: alignment); |
| 279 | |
| 280 | // Function aligned_alloc requires size to be a multiple of alignment; we |
| 281 | // pad the size to the next multiple if necessary. |
| 282 | if (!isMemRefSizeMultipleOf(type: memRefType, factor: alignment, op, defaultLayout: &defaultLayout)) |
| 283 | sizeBytes = createAligned(rewriter, loc, input: sizeBytes, alignment: allocAlignment); |
| 284 | |
| 285 | Type elementPtrType = this->getElementPtrType(type: memRefType); |
| 286 | auto results = |
| 287 | LLVM::CallOp::create(builder&: rewriter, location: loc, func: allocFuncOp.value(), |
| 288 | args: ValueRange({allocAlignment, sizeBytes})); |
| 289 | |
| 290 | Value ptr = |
| 291 | castAllocFuncResult(rewriter, loc, allocatedPtr: results.getResult(), memRefType, |
| 292 | elementPtrType, typeConverter: *getTypeConverter()); |
| 293 | |
| 294 | // Create the MemRef descriptor. |
| 295 | auto memRefDescriptor = this->createMemRefDescriptor( |
| 296 | loc, memRefType, allocatedPtr: ptr, alignedPtr: ptr, sizes, strides, rewriter); |
| 297 | |
| 298 | // Return the final value of the descriptor. |
| 299 | rewriter.replaceOp(op, newValues: {memRefDescriptor}); |
| 300 | return success(); |
| 301 | } |
| 302 | |
| 303 | /// The minimum alignment to use with aligned_alloc (has to be a power of 2). |
| 304 | static constexpr uint64_t kMinAlignedAllocAlignment = 16UL; |
| 305 | |
| 306 | /// Computes the alignment for aligned_alloc used to allocate the buffer for |
| 307 | /// the memory allocation op. |
| 308 | /// |
| 309 | /// Aligned_alloc requires the allocation size to be a power of two, and the |
| 310 | /// allocation size to be a multiple of the alignment. |
| 311 | int64_t alignedAllocationGetAlignment(memref::AllocOp op, |
| 312 | const DataLayout *defaultLayout) const { |
| 313 | if (std::optional<uint64_t> alignment = op.getAlignment()) |
| 314 | return *alignment; |
| 315 | |
| 316 | // Whenever we don't have alignment set, we will use an alignment |
| 317 | // consistent with the element type; since the allocation size has to be a |
| 318 | // power of two, we will bump to the next power of two if it isn't. |
| 319 | unsigned eltSizeBytes = getMemRefEltSizeInBytes( |
| 320 | typeConverter: getTypeConverter(), memRefType: op.getType(), op, defaultLayout); |
| 321 | return std::max(a: kMinAlignedAllocAlignment, |
| 322 | b: llvm::PowerOf2Ceil(A: eltSizeBytes)); |
| 323 | } |
| 324 | |
| 325 | /// Returns true if the memref size in bytes is known to be a multiple of |
| 326 | /// factor. |
| 327 | bool isMemRefSizeMultipleOf(MemRefType type, uint64_t factor, Operation *op, |
| 328 | const DataLayout *defaultLayout) const { |
| 329 | uint64_t sizeDivisor = |
| 330 | getMemRefEltSizeInBytes(typeConverter: getTypeConverter(), memRefType: type, op, defaultLayout); |
| 331 | for (unsigned i = 0, e = type.getRank(); i < e; i++) { |
| 332 | if (type.isDynamicDim(idx: i)) |
| 333 | continue; |
| 334 | sizeDivisor = sizeDivisor * type.getDimSize(idx: i); |
| 335 | } |
| 336 | return sizeDivisor % factor == 0; |
| 337 | } |
| 338 | |
| 339 | private: |
| 340 | /// Default layout to use in absence of the corresponding analysis. |
| 341 | DataLayout defaultLayout; |
| 342 | }; |
| 343 | |
| 344 | struct AllocaOpLowering : public ConvertOpToLLVMPattern<memref::AllocaOp> { |
| 345 | using ConvertOpToLLVMPattern<memref::AllocaOp>::ConvertOpToLLVMPattern; |
| 346 | |
| 347 | /// Allocates the underlying buffer using the right call. `allocatedBytePtr` |
| 348 | /// is set to null for stack allocations. `accessAlignment` is set if |
| 349 | /// alignment is needed post allocation (for eg. in conjunction with malloc). |
| 350 | LogicalResult |
| 351 | matchAndRewrite(memref::AllocaOp op, OpAdaptor adaptor, |
| 352 | ConversionPatternRewriter &rewriter) const override { |
| 353 | auto loc = op.getLoc(); |
| 354 | MemRefType memRefType = op.getType(); |
| 355 | if (!isConvertibleAndHasIdentityMaps(type: memRefType)) |
| 356 | return rewriter.notifyMatchFailure(arg&: op, msg: "incompatible memref type" ); |
| 357 | |
| 358 | // Get actual sizes of the memref as values: static sizes are constant |
| 359 | // values and dynamic sizes are passed to 'alloc' as operands. In case of |
| 360 | // zero-dimensional memref, assume a scalar (size 1). |
| 361 | SmallVector<Value, 4> sizes; |
| 362 | SmallVector<Value, 4> strides; |
| 363 | Value size; |
| 364 | |
| 365 | this->getMemRefDescriptorSizes(loc, memRefType, dynamicSizes: adaptor.getOperands(), |
| 366 | rewriter, sizes, strides, size, sizeInBytes: !true); |
| 367 | |
| 368 | // With alloca, one gets a pointer to the element type right away. |
| 369 | // For stack allocations. |
| 370 | auto elementType = |
| 371 | typeConverter->convertType(t: op.getType().getElementType()); |
| 372 | FailureOr<unsigned> maybeAddressSpace = |
| 373 | getTypeConverter()->getMemRefAddressSpace(type: op.getType()); |
| 374 | assert(succeeded(maybeAddressSpace) && "unsupported address space" ); |
| 375 | unsigned addrSpace = *maybeAddressSpace; |
| 376 | auto elementPtrType = |
| 377 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: addrSpace); |
| 378 | |
| 379 | auto allocatedElementPtr = |
| 380 | LLVM::AllocaOp::create(builder&: rewriter, location: loc, resultType: elementPtrType, elementType, arraySize: size, |
| 381 | alignment: op.getAlignment().value_or(u: 0)); |
| 382 | |
| 383 | // Create the MemRef descriptor. |
| 384 | auto memRefDescriptor = this->createMemRefDescriptor( |
| 385 | loc, memRefType, allocatedPtr: allocatedElementPtr, alignedPtr: allocatedElementPtr, sizes, |
| 386 | strides, rewriter); |
| 387 | |
| 388 | // Return the final value of the descriptor. |
| 389 | rewriter.replaceOp(op, newValues: {memRefDescriptor}); |
| 390 | return success(); |
| 391 | } |
| 392 | }; |
| 393 | |
| 394 | struct AllocaScopeOpLowering |
| 395 | : public ConvertOpToLLVMPattern<memref::AllocaScopeOp> { |
| 396 | using ConvertOpToLLVMPattern<memref::AllocaScopeOp>::ConvertOpToLLVMPattern; |
| 397 | |
| 398 | LogicalResult |
| 399 | matchAndRewrite(memref::AllocaScopeOp allocaScopeOp, OpAdaptor adaptor, |
| 400 | ConversionPatternRewriter &rewriter) const override { |
| 401 | OpBuilder::InsertionGuard guard(rewriter); |
| 402 | Location loc = allocaScopeOp.getLoc(); |
| 403 | |
| 404 | // Split the current block before the AllocaScopeOp to create the inlining |
| 405 | // point. |
| 406 | auto *currentBlock = rewriter.getInsertionBlock(); |
| 407 | auto *remainingOpsBlock = |
| 408 | rewriter.splitBlock(block: currentBlock, before: rewriter.getInsertionPoint()); |
| 409 | Block *continueBlock; |
| 410 | if (allocaScopeOp.getNumResults() == 0) { |
| 411 | continueBlock = remainingOpsBlock; |
| 412 | } else { |
| 413 | continueBlock = rewriter.createBlock( |
| 414 | insertBefore: remainingOpsBlock, argTypes: allocaScopeOp.getResultTypes(), |
| 415 | locs: SmallVector<Location>(allocaScopeOp->getNumResults(), |
| 416 | allocaScopeOp.getLoc())); |
| 417 | LLVM::BrOp::create(builder&: rewriter, location: loc, operands: ValueRange(), dest: remainingOpsBlock); |
| 418 | } |
| 419 | |
| 420 | // Inline body region. |
| 421 | Block *beforeBody = &allocaScopeOp.getBodyRegion().front(); |
| 422 | Block *afterBody = &allocaScopeOp.getBodyRegion().back(); |
| 423 | rewriter.inlineRegionBefore(region&: allocaScopeOp.getBodyRegion(), before: continueBlock); |
| 424 | |
| 425 | // Save stack and then branch into the body of the region. |
| 426 | rewriter.setInsertionPointToEnd(currentBlock); |
| 427 | auto stackSaveOp = LLVM::StackSaveOp::create(builder&: rewriter, location: loc, res: getPtrType()); |
| 428 | LLVM::BrOp::create(builder&: rewriter, location: loc, operands: ValueRange(), dest: beforeBody); |
| 429 | |
| 430 | // Replace the alloca_scope return with a branch that jumps out of the body. |
| 431 | // Stack restore before leaving the body region. |
| 432 | rewriter.setInsertionPointToEnd(afterBody); |
| 433 | auto returnOp = |
| 434 | cast<memref::AllocaScopeReturnOp>(Val: afterBody->getTerminator()); |
| 435 | auto branchOp = rewriter.replaceOpWithNewOp<LLVM::BrOp>( |
| 436 | op: returnOp, args: returnOp.getResults(), args&: continueBlock); |
| 437 | |
| 438 | // Insert stack restore before jumping out the body of the region. |
| 439 | rewriter.setInsertionPoint(branchOp); |
| 440 | LLVM::StackRestoreOp::create(builder&: rewriter, location: loc, ptr: stackSaveOp); |
| 441 | |
| 442 | // Replace the op with values return from the body region. |
| 443 | rewriter.replaceOp(op: allocaScopeOp, newValues: continueBlock->getArguments()); |
| 444 | |
| 445 | return success(); |
| 446 | } |
| 447 | }; |
| 448 | |
| 449 | struct AssumeAlignmentOpLowering |
| 450 | : public ConvertOpToLLVMPattern<memref::AssumeAlignmentOp> { |
| 451 | using ConvertOpToLLVMPattern< |
| 452 | memref::AssumeAlignmentOp>::ConvertOpToLLVMPattern; |
| 453 | explicit AssumeAlignmentOpLowering(const LLVMTypeConverter &converter) |
| 454 | : ConvertOpToLLVMPattern<memref::AssumeAlignmentOp>(converter) {} |
| 455 | |
| 456 | LogicalResult |
| 457 | matchAndRewrite(memref::AssumeAlignmentOp op, OpAdaptor adaptor, |
| 458 | ConversionPatternRewriter &rewriter) const override { |
| 459 | Value memref = adaptor.getMemref(); |
| 460 | unsigned alignment = op.getAlignment(); |
| 461 | auto loc = op.getLoc(); |
| 462 | |
| 463 | auto srcMemRefType = cast<MemRefType>(Val: op.getMemref().getType()); |
| 464 | Value ptr = getStridedElementPtr(rewriter, loc, type: srcMemRefType, memRefDesc: memref, |
| 465 | /*indices=*/{}); |
| 466 | |
| 467 | // Emit llvm.assume(true) ["align"(memref, alignment)]. |
| 468 | // This is more direct than ptrtoint-based checks, is explicitly supported, |
| 469 | // and works with non-integral address spaces. |
| 470 | Value trueCond = |
| 471 | LLVM::ConstantOp::create(builder&: rewriter, location: loc, value: rewriter.getBoolAttr(value: true)); |
| 472 | Value alignmentConst = |
| 473 | createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: alignment); |
| 474 | LLVM::AssumeOp::create(builder&: rewriter, location: loc, cond: trueCond, tag: LLVM::AssumeAlignTag(), ptr, |
| 475 | align: alignmentConst); |
| 476 | rewriter.replaceOp(op, newValues: memref); |
| 477 | return success(); |
| 478 | } |
| 479 | }; |
| 480 | |
| 481 | struct DistinctObjectsOpLowering |
| 482 | : public ConvertOpToLLVMPattern<memref::DistinctObjectsOp> { |
| 483 | using ConvertOpToLLVMPattern< |
| 484 | memref::DistinctObjectsOp>::ConvertOpToLLVMPattern; |
| 485 | explicit DistinctObjectsOpLowering(const LLVMTypeConverter &converter) |
| 486 | : ConvertOpToLLVMPattern<memref::DistinctObjectsOp>(converter) {} |
| 487 | |
| 488 | LogicalResult |
| 489 | matchAndRewrite(memref::DistinctObjectsOp op, OpAdaptor adaptor, |
| 490 | ConversionPatternRewriter &rewriter) const override { |
| 491 | ValueRange operands = adaptor.getOperands(); |
| 492 | if (operands.size() <= 1) { |
| 493 | // Fast path. |
| 494 | rewriter.replaceOp(op, newValues: operands); |
| 495 | return success(); |
| 496 | } |
| 497 | |
| 498 | Location loc = op.getLoc(); |
| 499 | SmallVector<Value> ptrs; |
| 500 | for (auto [origOperand, newOperand] : |
| 501 | llvm::zip_equal(t: op.getOperands(), u&: operands)) { |
| 502 | auto memrefType = cast<MemRefType>(Val: origOperand.getType()); |
| 503 | MemRefDescriptor memRefDescriptor(newOperand); |
| 504 | Value ptr = memRefDescriptor.bufferPtr(builder&: rewriter, loc, converter: *getTypeConverter(), |
| 505 | type: memrefType); |
| 506 | ptrs.push_back(Elt: ptr); |
| 507 | } |
| 508 | |
| 509 | auto cond = |
| 510 | LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: rewriter.getI1Type(), value: 1); |
| 511 | // Generate separate_storage assumptions for each pair of pointers. |
| 512 | for (auto i : llvm::seq<size_t>(Size: ptrs.size() - 1)) { |
| 513 | for (auto j : llvm::seq<size_t>(Begin: i + 1, End: ptrs.size())) { |
| 514 | Value ptr1 = ptrs[i]; |
| 515 | Value ptr2 = ptrs[j]; |
| 516 | LLVM::AssumeOp::create(builder&: rewriter, location: loc, cond, |
| 517 | tag: LLVM::AssumeSeparateStorageTag{}, ptr1, ptr2); |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | rewriter.replaceOp(op, newValues: operands); |
| 522 | return success(); |
| 523 | } |
| 524 | }; |
| 525 | |
| 526 | // A `dealloc` is converted into a call to `free` on the underlying data buffer. |
| 527 | // The memref descriptor being an SSA value, there is no need to clean it up |
| 528 | // in any way. |
| 529 | class DeallocOpLowering : public ConvertOpToLLVMPattern<memref::DeallocOp> { |
| 530 | SymbolTableCollection *symbolTables = nullptr; |
| 531 | |
| 532 | public: |
| 533 | explicit DeallocOpLowering(const LLVMTypeConverter &typeConverter, |
| 534 | SymbolTableCollection *symbolTables = nullptr, |
| 535 | PatternBenefit benefit = 1) |
| 536 | : ConvertOpToLLVMPattern<memref::DeallocOp>(typeConverter, benefit), |
| 537 | symbolTables(symbolTables) {} |
| 538 | |
| 539 | LogicalResult |
| 540 | matchAndRewrite(memref::DeallocOp op, OpAdaptor adaptor, |
| 541 | ConversionPatternRewriter &rewriter) const override { |
| 542 | // Insert the `free` declaration if it is not already present. |
| 543 | FailureOr<LLVM::LLVMFuncOp> freeFunc = |
| 544 | getFreeFn(b&: rewriter, typeConverter: getTypeConverter(), |
| 545 | module: op->getParentWithTrait<OpTrait::SymbolTable>(), symbolTables); |
| 546 | if (failed(Result: freeFunc)) |
| 547 | return failure(); |
| 548 | Value allocatedPtr; |
| 549 | if (auto unrankedTy = |
| 550 | llvm::dyn_cast<UnrankedMemRefType>(Val: op.getMemref().getType())) { |
| 551 | auto elementPtrTy = LLVM::LLVMPointerType::get( |
| 552 | context: rewriter.getContext(), addressSpace: unrankedTy.getMemorySpaceAsInt()); |
| 553 | allocatedPtr = UnrankedMemRefDescriptor::allocatedPtr( |
| 554 | builder&: rewriter, loc: op.getLoc(), |
| 555 | memRefDescPtr: UnrankedMemRefDescriptor(adaptor.getMemref()) |
| 556 | .memRefDescPtr(builder&: rewriter, loc: op.getLoc()), |
| 557 | elemPtrType: elementPtrTy); |
| 558 | } else { |
| 559 | allocatedPtr = MemRefDescriptor(adaptor.getMemref()) |
| 560 | .allocatedPtr(builder&: rewriter, loc: op.getLoc()); |
| 561 | } |
| 562 | // `free` takes a pointer in the default address space. Undo the cast that |
| 563 | // the alloc lowering inserts for memrefs in another address space. |
| 564 | Type freeArgTy = getPtrType(); |
| 565 | if (allocatedPtr.getType() != freeArgTy) |
| 566 | allocatedPtr = LLVM::AddrSpaceCastOp::create(builder&: rewriter, location: op.getLoc(), |
| 567 | res: freeArgTy, arg: allocatedPtr); |
| 568 | rewriter.replaceOpWithNewOp<LLVM::CallOp>(op, args&: freeFunc.value(), |
| 569 | args&: allocatedPtr); |
| 570 | return success(); |
| 571 | } |
| 572 | }; |
| 573 | |
| 574 | // A `dim` is converted to a constant for static sizes and to an access to the |
| 575 | // size stored in the memref descriptor for dynamic sizes. |
| 576 | struct DimOpLowering : public ConvertOpToLLVMPattern<memref::DimOp> { |
| 577 | using ConvertOpToLLVMPattern<memref::DimOp>::ConvertOpToLLVMPattern; |
| 578 | |
| 579 | LogicalResult |
| 580 | matchAndRewrite(memref::DimOp dimOp, OpAdaptor adaptor, |
| 581 | ConversionPatternRewriter &rewriter) const override { |
| 582 | Type operandType = dimOp.getSource().getType(); |
| 583 | if (isa<UnrankedMemRefType>(Val: operandType)) { |
| 584 | FailureOr<Value> = extractSizeOfUnrankedMemRef( |
| 585 | operandType, dimOp, adaptor: adaptor.getOperands(), rewriter); |
| 586 | if (failed(Result: extractedSize)) |
| 587 | return failure(); |
| 588 | rewriter.replaceOp(op: dimOp, newValues: {*extractedSize}); |
| 589 | return success(); |
| 590 | } |
| 591 | if (isa<MemRefType>(Val: operandType)) { |
| 592 | rewriter.replaceOp( |
| 593 | op: dimOp, newValues: {extractSizeOfRankedMemRef(operandType, dimOp, |
| 594 | adaptor: adaptor.getOperands(), rewriter)}); |
| 595 | return success(); |
| 596 | } |
| 597 | llvm_unreachable("expected MemRefType or UnrankedMemRefType" ); |
| 598 | } |
| 599 | |
| 600 | private: |
| 601 | FailureOr<Value> |
| 602 | (Type operandType, memref::DimOp dimOp, |
| 603 | OpAdaptor adaptor, |
| 604 | ConversionPatternRewriter &rewriter) const { |
| 605 | Location loc = dimOp.getLoc(); |
| 606 | |
| 607 | auto unrankedMemRefType = cast<UnrankedMemRefType>(Val&: operandType); |
| 608 | auto scalarMemRefType = |
| 609 | MemRefType::get(shape: {}, elementType: unrankedMemRefType.getElementType()); |
| 610 | FailureOr<unsigned> maybeAddressSpace = |
| 611 | getTypeConverter()->getMemRefAddressSpace(type: unrankedMemRefType); |
| 612 | if (failed(Result: maybeAddressSpace)) { |
| 613 | dimOp.emitOpError(message: "memref memory space must be convertible to an integer " |
| 614 | "address space" ); |
| 615 | return failure(); |
| 616 | } |
| 617 | unsigned addressSpace = *maybeAddressSpace; |
| 618 | |
| 619 | // Extract pointer to the underlying ranked descriptor and bitcast it to a |
| 620 | // memref<element_type> descriptor pointer to minimize the number of GEP |
| 621 | // operations. |
| 622 | UnrankedMemRefDescriptor unrankedDesc(adaptor.getSource()); |
| 623 | Value underlyingRankedDesc = unrankedDesc.memRefDescPtr(builder&: rewriter, loc); |
| 624 | |
| 625 | Type elementType = typeConverter->convertType(t: scalarMemRefType); |
| 626 | |
| 627 | // Get pointer to offset field of memref<element_type> descriptor. |
| 628 | auto indexPtrTy = |
| 629 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace); |
| 630 | Value offsetPtr = |
| 631 | LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: indexPtrTy, elementType, |
| 632 | basePtr: underlyingRankedDesc, indices: ArrayRef<LLVM::GEPArg>{0, 2}); |
| 633 | |
| 634 | // The size value that we have to extract can be obtained using GEPop with |
| 635 | // `dimOp.index() + 1` index argument. |
| 636 | Value idxPlusOne = LLVM::AddOp::create( |
| 637 | builder&: rewriter, location: loc, |
| 638 | lhs: createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: 1), |
| 639 | rhs: adaptor.getIndex()); |
| 640 | Value sizePtr = LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: indexPtrTy, |
| 641 | elementType: getTypeConverter()->getIndexType(), |
| 642 | basePtr: offsetPtr, indices: idxPlusOne); |
| 643 | return LLVM::LoadOp::create(builder&: rewriter, location: loc, |
| 644 | type: getTypeConverter()->getIndexType(), addr: sizePtr) |
| 645 | .getResult(); |
| 646 | } |
| 647 | |
| 648 | std::optional<int64_t> getConstantDimIndex(memref::DimOp dimOp) const { |
| 649 | if (auto idx = dimOp.getConstantIndex()) |
| 650 | return idx; |
| 651 | |
| 652 | if (auto constantOp = dimOp.getIndex().getDefiningOp<LLVM::ConstantOp>()) |
| 653 | return cast<IntegerAttr>(Val: constantOp.getValue()).getValue().getSExtValue(); |
| 654 | |
| 655 | return std::nullopt; |
| 656 | } |
| 657 | |
| 658 | Value (Type operandType, memref::DimOp dimOp, |
| 659 | OpAdaptor adaptor, |
| 660 | ConversionPatternRewriter &rewriter) const { |
| 661 | Location loc = dimOp.getLoc(); |
| 662 | |
| 663 | // Take advantage if index is constant. |
| 664 | MemRefType memRefType = cast<MemRefType>(Val&: operandType); |
| 665 | Type indexType = getIndexType(); |
| 666 | if (std::optional<int64_t> index = getConstantDimIndex(dimOp)) { |
| 667 | int64_t i = *index; |
| 668 | if (i >= 0 && i < memRefType.getRank()) { |
| 669 | if (memRefType.isDynamicDim(idx: i)) { |
| 670 | // extract dynamic size from the memref descriptor. |
| 671 | MemRefDescriptor descriptor(adaptor.getSource()); |
| 672 | return descriptor.size(builder&: rewriter, loc, pos: i); |
| 673 | } |
| 674 | // Use constant for static size. |
| 675 | int64_t dimSize = memRefType.getDimSize(idx: i); |
| 676 | return createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: dimSize); |
| 677 | } |
| 678 | } |
| 679 | Value index = adaptor.getIndex(); |
| 680 | int64_t rank = memRefType.getRank(); |
| 681 | MemRefDescriptor memrefDescriptor(adaptor.getSource()); |
| 682 | return memrefDescriptor.size(builder&: rewriter, loc, pos: index, rank); |
| 683 | } |
| 684 | }; |
| 685 | |
| 686 | /// Common base for load and store operations on MemRefs. Restricts the match |
| 687 | /// to supported MemRef types. Provides functionality to emit code accessing a |
| 688 | /// specific element of the underlying data buffer. |
| 689 | template <typename Derived> |
| 690 | struct LoadStoreOpLowering : public ConvertOpToLLVMPattern<Derived> { |
| 691 | using ConvertOpToLLVMPattern<Derived>::ConvertOpToLLVMPattern; |
| 692 | using ConvertOpToLLVMPattern<Derived>::isConvertibleAndHasIdentityMaps; |
| 693 | using Base = LoadStoreOpLowering<Derived>; |
| 694 | }; |
| 695 | |
| 696 | /// Wrap a llvm.cmpxchg operation in a while loop so that the operation can be |
| 697 | /// retried until it succeeds in atomically storing a new value into memory. |
| 698 | /// |
| 699 | /// +---------------------------------+ |
| 700 | /// | <code before the AtomicRMWOp> | |
| 701 | /// | <compute initial %loaded> | |
| 702 | /// | cf.br loop(%loaded) | |
| 703 | /// +---------------------------------+ |
| 704 | /// | |
| 705 | /// -------| | |
| 706 | /// | v v |
| 707 | /// | +--------------------------------+ |
| 708 | /// | | loop(%loaded): | |
| 709 | /// | | <body contents> | |
| 710 | /// | | %pair = cmpxchg | |
| 711 | /// | | %ok = %pair[0] | |
| 712 | /// | | %new = %pair[1] | |
| 713 | /// | | cf.cond_br %ok, end, loop(%new) | |
| 714 | /// | +--------------------------------+ |
| 715 | /// | | | |
| 716 | /// |----------- | |
| 717 | /// v |
| 718 | /// +--------------------------------+ |
| 719 | /// | end: | |
| 720 | /// | <code after the AtomicRMWOp> | |
| 721 | /// +--------------------------------+ |
| 722 | /// |
| 723 | struct GenericAtomicRMWOpLowering |
| 724 | : public LoadStoreOpLowering<memref::GenericAtomicRMWOp> { |
| 725 | using Base::Base; |
| 726 | |
| 727 | LogicalResult |
| 728 | matchAndRewrite(memref::GenericAtomicRMWOp atomicOp, OpAdaptor adaptor, |
| 729 | ConversionPatternRewriter &rewriter) const override { |
| 730 | auto loc = atomicOp.getLoc(); |
| 731 | Type valueType = typeConverter->convertType(t: atomicOp.getResult().getType()); |
| 732 | |
| 733 | // `llvm.cmpxchg` only supports integer or pointer operands. For |
| 734 | // floating-point element types, perform the CAS on a same-width integer |
| 735 | // and bitcast at the boundaries. |
| 736 | bool needsBitcast = isa<FloatType>(Val: valueType); |
| 737 | Type cmpxchgType = valueType; |
| 738 | if (needsBitcast) { |
| 739 | unsigned bitWidth = cast<FloatType>(Val&: valueType).getWidth(); |
| 740 | cmpxchgType = rewriter.getIntegerType(width: bitWidth); |
| 741 | } |
| 742 | |
| 743 | // Split the block into initial, loop, and ending parts. |
| 744 | auto *initBlock = rewriter.getInsertionBlock(); |
| 745 | auto *loopBlock = rewriter.splitBlock(block: initBlock, before: Block::iterator(atomicOp)); |
| 746 | loopBlock->addArgument(type: cmpxchgType, loc); |
| 747 | |
| 748 | auto *endBlock = |
| 749 | rewriter.splitBlock(block: loopBlock, before: Block::iterator(atomicOp)++); |
| 750 | |
| 751 | // Compute the loaded value and branch to the loop block. |
| 752 | rewriter.setInsertionPointToEnd(initBlock); |
| 753 | auto memRefType = cast<MemRefType>(Val: atomicOp.getMemref().getType()); |
| 754 | auto dataPtr = getStridedElementPtr( |
| 755 | rewriter, loc, type: memRefType, memRefDesc: adaptor.getMemref(), indices: adaptor.getIndices()); |
| 756 | Value init = LLVM::LoadOp::create( |
| 757 | builder&: rewriter, location: loc, type: typeConverter->convertType(t: memRefType.getElementType()), |
| 758 | addr: dataPtr); |
| 759 | if (needsBitcast) |
| 760 | init = LLVM::BitcastOp::create(builder&: rewriter, location: loc, res: cmpxchgType, arg: init); |
| 761 | LLVM::BrOp::create(builder&: rewriter, location: loc, operands: init, dest: loopBlock); |
| 762 | |
| 763 | // Prepare the body of the loop block. |
| 764 | rewriter.setInsertionPointToStart(loopBlock); |
| 765 | |
| 766 | // Clone the GenericAtomicRMWOp region and extract the result. |
| 767 | Value loopArgument = loopBlock->getArgument(i: 0); |
| 768 | Value loopArgForBody = loopArgument; |
| 769 | if (needsBitcast) |
| 770 | loopArgForBody = |
| 771 | LLVM::BitcastOp::create(builder&: rewriter, location: loc, res: valueType, arg: loopArgument); |
| 772 | IRMapping mapping; |
| 773 | mapping.map(from: atomicOp.getCurrentValue(), to: loopArgForBody); |
| 774 | Block &entryBlock = atomicOp.body().front(); |
| 775 | for (auto &nestedOp : entryBlock.without_terminator()) { |
| 776 | Operation *clone = rewriter.clone(op&: nestedOp, mapper&: mapping); |
| 777 | mapping.map(from: nestedOp.getResults(), to: clone->getResults()); |
| 778 | } |
| 779 | |
| 780 | // Values defined outside the region are not in the mapping. |
| 781 | Value result = |
| 782 | mapping.lookupOrDefault(from: entryBlock.getTerminator()->getOperand(idx: 0)); |
| 783 | if (needsBitcast) |
| 784 | result = LLVM::BitcastOp::create(builder&: rewriter, location: loc, res: cmpxchgType, arg: result); |
| 785 | |
| 786 | // Prepare the epilog of the loop block. |
| 787 | // Append the cmpxchg op to the end of the loop block. |
| 788 | auto successOrdering = LLVM::AtomicOrdering::acq_rel; |
| 789 | auto failureOrdering = LLVM::AtomicOrdering::monotonic; |
| 790 | auto cmpxchg = |
| 791 | LLVM::AtomicCmpXchgOp::create(builder&: rewriter, location: loc, ptr: dataPtr, cmp: loopArgument, |
| 792 | val: result, successOrdering, failureOrdering); |
| 793 | // Extract the %new_loaded and %ok values from the pair. |
| 794 | Value newLoaded = LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: cmpxchg, position: 0); |
| 795 | Value ok = LLVM::ExtractValueOp::create(builder&: rewriter, location: loc, container: cmpxchg, position: 1); |
| 796 | |
| 797 | // Conditionally branch to the end or back to the loop depending on %ok. |
| 798 | LLVM::CondBrOp::create(builder&: rewriter, location: loc, condition: ok, trueDest: endBlock, trueOperands: ArrayRef<Value>(), |
| 799 | falseDest: loopBlock, falseOperands: newLoaded); |
| 800 | |
| 801 | // The 'result' of the atomic_rmw op is the newly loaded value. Bitcast |
| 802 | // back to the float type if needed. Insert at the start of `endBlock` so |
| 803 | // the bitcast precedes the existing terminator (split into endBlock). |
| 804 | if (needsBitcast) { |
| 805 | rewriter.setInsertionPointToStart(endBlock); |
| 806 | newLoaded = LLVM::BitcastOp::create(builder&: rewriter, location: loc, res: valueType, arg: newLoaded); |
| 807 | } |
| 808 | rewriter.setInsertionPointToEnd(endBlock); |
| 809 | rewriter.replaceOp(op: atomicOp, newValues: {newLoaded}); |
| 810 | |
| 811 | return success(); |
| 812 | } |
| 813 | }; |
| 814 | |
| 815 | /// Returns the LLVM type of the global variable given the memref type `type`. |
| 816 | static Type |
| 817 | convertGlobalMemrefTypeToLLVM(MemRefType type, |
| 818 | const LLVMTypeConverter &typeConverter) { |
| 819 | // LLVM type for a global memref will be a multi-dimension array. For |
| 820 | // declarations or uninitialized global memrefs, we can potentially flatten |
| 821 | // this to a 1D array. However, for memref.global's with an initial value, |
| 822 | // we do not intend to flatten the ElementsAttribute when going from std -> |
| 823 | // LLVM dialect, so the LLVM type needs to me a multi-dimension array. |
| 824 | Type elementType = typeConverter.convertType(t: type.getElementType()); |
| 825 | Type arrayTy = elementType; |
| 826 | // Shape has the outermost dim at index 0, so need to walk it backwards |
| 827 | for (int64_t dim : llvm::reverse(C: type.getShape())) |
| 828 | arrayTy = LLVM::LLVMArrayType::get(elementType: arrayTy, numElements: dim); |
| 829 | return arrayTy; |
| 830 | } |
| 831 | |
| 832 | /// GlobalMemrefOp is lowered to a LLVM Global Variable. |
| 833 | class GlobalMemrefOpLowering : public ConvertOpToLLVMPattern<memref::GlobalOp> { |
| 834 | SymbolTableCollection *symbolTables = nullptr; |
| 835 | |
| 836 | public: |
| 837 | explicit GlobalMemrefOpLowering(const LLVMTypeConverter &typeConverter, |
| 838 | SymbolTableCollection *symbolTables = nullptr, |
| 839 | PatternBenefit benefit = 1) |
| 840 | : ConvertOpToLLVMPattern<memref::GlobalOp>(typeConverter, benefit), |
| 841 | symbolTables(symbolTables) {} |
| 842 | |
| 843 | LogicalResult |
| 844 | matchAndRewrite(memref::GlobalOp global, OpAdaptor adaptor, |
| 845 | ConversionPatternRewriter &rewriter) const override { |
| 846 | MemRefType type = global.getType(); |
| 847 | if (!isConvertibleAndHasIdentityMaps(type)) |
| 848 | return failure(); |
| 849 | |
| 850 | Type arrayTy = convertGlobalMemrefTypeToLLVM(type, typeConverter: *getTypeConverter()); |
| 851 | |
| 852 | LLVM::Linkage linkage = |
| 853 | global.isPublic() ? LLVM::Linkage::External : LLVM::Linkage::Private; |
| 854 | bool isExternal = global.isExternal(); |
| 855 | bool isUninitialized = global.isUninitialized(); |
| 856 | |
| 857 | Attribute initialValue = nullptr; |
| 858 | if (!isExternal && !isUninitialized) { |
| 859 | auto elementsAttr = llvm::cast<ElementsAttr>(Val: *global.getInitialValue()); |
| 860 | initialValue = elementsAttr; |
| 861 | |
| 862 | // For scalar memrefs, the global variable created is of the element type, |
| 863 | // so unpack the elements attribute to extract the value. |
| 864 | if (type.getRank() == 0) |
| 865 | initialValue = elementsAttr.getSplatValue<Attribute>(); |
| 866 | } |
| 867 | |
| 868 | uint64_t alignment = global.getAlignment().value_or(u: 0); |
| 869 | FailureOr<unsigned> addressSpace = |
| 870 | getTypeConverter()->getMemRefAddressSpace(type); |
| 871 | if (failed(Result: addressSpace)) |
| 872 | return global.emitOpError( |
| 873 | message: "memory space cannot be converted to an integer address space" ); |
| 874 | |
| 875 | // Remove old operation from symbol table. |
| 876 | SymbolTable *symbolTable = nullptr; |
| 877 | if (symbolTables) { |
| 878 | Operation *symbolTableOp = |
| 879 | global->getParentWithTrait<OpTrait::SymbolTable>(); |
| 880 | symbolTable = &symbolTables->getSymbolTable(op: symbolTableOp); |
| 881 | symbolTable->remove(op: global); |
| 882 | } |
| 883 | |
| 884 | // Create new operation. |
| 885 | auto newGlobal = rewriter.replaceOpWithNewOp<LLVM::GlobalOp>( |
| 886 | op: global, args&: arrayTy, args: global.getConstant(), args&: linkage, args: global.getSymName(), |
| 887 | args&: initialValue, args&: alignment, args&: *addressSpace); |
| 888 | |
| 889 | // Insert new operation into symbol table. |
| 890 | if (symbolTable) |
| 891 | symbolTable->insert(symbol: newGlobal, insertPt: rewriter.getInsertionPoint()); |
| 892 | |
| 893 | if (!isExternal && isUninitialized) { |
| 894 | rewriter.createBlock(parent: &newGlobal.getInitializerRegion()); |
| 895 | Value undef[] = { |
| 896 | LLVM::UndefOp::create(builder&: rewriter, location: newGlobal.getLoc(), res: arrayTy)}; |
| 897 | LLVM::ReturnOp::create(builder&: rewriter, location: newGlobal.getLoc(), args: undef); |
| 898 | } |
| 899 | return success(); |
| 900 | } |
| 901 | }; |
| 902 | |
| 903 | /// GetGlobalMemrefOp is lowered into a Memref descriptor with the pointer to |
| 904 | /// the first element stashed into the descriptor. This reuses |
| 905 | /// `AllocLikeOpLowering` to reuse the Memref descriptor construction. |
| 906 | struct GetGlobalMemrefOpLowering |
| 907 | : public ConvertOpToLLVMPattern<memref::GetGlobalOp> { |
| 908 | using ConvertOpToLLVMPattern<memref::GetGlobalOp>::ConvertOpToLLVMPattern; |
| 909 | |
| 910 | /// Buffer "allocation" for memref.get_global op is getting the address of |
| 911 | /// the global variable referenced. |
| 912 | LogicalResult |
| 913 | matchAndRewrite(memref::GetGlobalOp op, OpAdaptor adaptor, |
| 914 | ConversionPatternRewriter &rewriter) const override { |
| 915 | auto loc = op.getLoc(); |
| 916 | MemRefType memRefType = op.getType(); |
| 917 | if (!isConvertibleAndHasIdentityMaps(type: memRefType)) |
| 918 | return rewriter.notifyMatchFailure(arg&: op, msg: "incompatible memref type" ); |
| 919 | |
| 920 | // Get actual sizes of the memref as values: static sizes are constant |
| 921 | // values and dynamic sizes are passed to 'alloc' as operands. In case of |
| 922 | // zero-dimensional memref, assume a scalar (size 1). |
| 923 | SmallVector<Value, 4> sizes; |
| 924 | SmallVector<Value, 4> strides; |
| 925 | Value sizeBytes; |
| 926 | |
| 927 | this->getMemRefDescriptorSizes(loc, memRefType, dynamicSizes: adaptor.getOperands(), |
| 928 | rewriter, sizes, strides, size&: sizeBytes, sizeInBytes: !false); |
| 929 | |
| 930 | MemRefType type = cast<MemRefType>(Val: op.getResult().getType()); |
| 931 | |
| 932 | // This is called after a type conversion, which would have failed if this |
| 933 | // call fails. |
| 934 | FailureOr<unsigned> maybeAddressSpace = |
| 935 | getTypeConverter()->getMemRefAddressSpace(type); |
| 936 | assert(succeeded(maybeAddressSpace) && "unsupported address space" ); |
| 937 | unsigned memSpace = *maybeAddressSpace; |
| 938 | |
| 939 | Type arrayTy = convertGlobalMemrefTypeToLLVM(type, typeConverter: *getTypeConverter()); |
| 940 | auto ptrTy = LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: memSpace); |
| 941 | auto addressOf = |
| 942 | LLVM::AddressOfOp::create(builder&: rewriter, location: loc, res: ptrTy, global_name: op.getName()); |
| 943 | |
| 944 | // Get the address of the first element in the array by creating a GEP with |
| 945 | // the address of the GV as the base, and (rank + 1) number of 0 indices. |
| 946 | auto gep = |
| 947 | LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: ptrTy, elementType: arrayTy, basePtr: addressOf, |
| 948 | indices: SmallVector<LLVM::GEPArg>(type.getRank() + 1, 0)); |
| 949 | |
| 950 | // We do not expect the memref obtained using `memref.get_global` to be |
| 951 | // ever deallocated. Set the allocated pointer to be known bad value to |
| 952 | // help debug if that ever happens. |
| 953 | auto intPtrType = getIntPtrType(addressSpace: memSpace); |
| 954 | Value deadBeefConst = |
| 955 | createIndexAttrConstant(builder&: rewriter, loc: op->getLoc(), resultType: intPtrType, value: 0xdeadbeef); |
| 956 | auto deadBeefPtr = |
| 957 | LLVM::IntToPtrOp::create(builder&: rewriter, location: loc, resultType: ptrTy, operands: deadBeefConst); |
| 958 | |
| 959 | // Both allocated and aligned pointers are same. We could potentially stash |
| 960 | // a nullptr for the allocated pointer since we do not expect any dealloc. |
| 961 | // Create the MemRef descriptor. |
| 962 | auto memRefDescriptor = this->createMemRefDescriptor( |
| 963 | loc, memRefType, allocatedPtr: deadBeefPtr, alignedPtr: gep, sizes, strides, rewriter); |
| 964 | |
| 965 | // Return the final value of the descriptor. |
| 966 | rewriter.replaceOp(op, newValues: {memRefDescriptor}); |
| 967 | return success(); |
| 968 | } |
| 969 | }; |
| 970 | |
| 971 | // Load operation is lowered to obtaining a pointer to the indexed element |
| 972 | // and loading it. |
| 973 | struct LoadOpLowering : public LoadStoreOpLowering<memref::LoadOp> { |
| 974 | using Base::Base; |
| 975 | |
| 976 | LogicalResult |
| 977 | matchAndRewrite(memref::LoadOp loadOp, OpAdaptor adaptor, |
| 978 | ConversionPatternRewriter &rewriter) const override { |
| 979 | auto type = loadOp.getMemRefType(); |
| 980 | |
| 981 | // Per memref.load spec, the indices must be in-bounds: |
| 982 | // 0 <= idx < dim_size, and additionally all offsets are non-negative, |
| 983 | // hence inbounds and nuw are used when lowering to llvm.getelementptr. |
| 984 | Value dataPtr = getStridedElementPtr( |
| 985 | rewriter, loc: loadOp.getLoc(), type, memRefDesc: adaptor.getMemref(), |
| 986 | indices: adaptor.getIndices(), noWrapFlags: getLoadStoreNoWrapFlags(type)); |
| 987 | rewriter.replaceOpWithNewOp<LLVM::LoadOp>( |
| 988 | op: loadOp, args: typeConverter->convertType(t: type.getElementType()), args&: dataPtr, |
| 989 | args: loadOp.getAlignment().value_or(u: 0), args: false, args: loadOp.getNontemporal(), |
| 990 | /*isInvariant=*/args: loadOp.getInvariant()); |
| 991 | return success(); |
| 992 | } |
| 993 | }; |
| 994 | |
| 995 | // Store operation is lowered to obtaining a pointer to the indexed element, |
| 996 | // and storing the given value to it. |
| 997 | struct StoreOpLowering : public LoadStoreOpLowering<memref::StoreOp> { |
| 998 | using Base::Base; |
| 999 | |
| 1000 | LogicalResult |
| 1001 | matchAndRewrite(memref::StoreOp op, OpAdaptor adaptor, |
| 1002 | ConversionPatternRewriter &rewriter) const override { |
| 1003 | auto type = op.getMemRefType(); |
| 1004 | |
| 1005 | // Per memref.store spec, the indices must be in-bounds: |
| 1006 | // 0 <= idx < dim_size, and additionally all offsets are non-negative, |
| 1007 | // hence inbounds and nuw are used when lowering to llvm.getelementptr. |
| 1008 | Value dataPtr = getStridedElementPtr( |
| 1009 | rewriter, loc: op.getLoc(), type, memRefDesc: adaptor.getMemref(), indices: adaptor.getIndices(), |
| 1010 | noWrapFlags: getLoadStoreNoWrapFlags(type)); |
| 1011 | rewriter.replaceOpWithNewOp<LLVM::StoreOp>(op, args: adaptor.getValue(), args&: dataPtr, |
| 1012 | args: op.getAlignment().value_or(u: 0), |
| 1013 | args: false, args: op.getNontemporal()); |
| 1014 | return success(); |
| 1015 | } |
| 1016 | }; |
| 1017 | |
| 1018 | // The prefetch operation is lowered in a way similar to the load operation |
| 1019 | // except that the llvm.prefetch operation is used for replacement. |
| 1020 | struct PrefetchOpLowering : public LoadStoreOpLowering<memref::PrefetchOp> { |
| 1021 | using Base::Base; |
| 1022 | |
| 1023 | LogicalResult |
| 1024 | matchAndRewrite(memref::PrefetchOp prefetchOp, OpAdaptor adaptor, |
| 1025 | ConversionPatternRewriter &rewriter) const override { |
| 1026 | auto type = prefetchOp.getMemRefType(); |
| 1027 | auto loc = prefetchOp.getLoc(); |
| 1028 | |
| 1029 | Value dataPtr = getStridedElementPtr( |
| 1030 | rewriter, loc, type, memRefDesc: adaptor.getMemref(), indices: adaptor.getIndices()); |
| 1031 | |
| 1032 | // Replace with llvm.prefetch. |
| 1033 | IntegerAttr isWrite = rewriter.getI32IntegerAttr(value: prefetchOp.getIsWrite()); |
| 1034 | IntegerAttr localityHint = prefetchOp.getLocalityHintAttr(); |
| 1035 | IntegerAttr isData = |
| 1036 | rewriter.getI32IntegerAttr(value: prefetchOp.getIsDataCache()); |
| 1037 | rewriter.replaceOpWithNewOp<LLVM::Prefetch>(op: prefetchOp, args&: dataPtr, args&: isWrite, |
| 1038 | args&: localityHint, args&: isData); |
| 1039 | return success(); |
| 1040 | } |
| 1041 | }; |
| 1042 | |
| 1043 | struct RankOpLowering : public ConvertOpToLLVMPattern<memref::RankOp> { |
| 1044 | using ConvertOpToLLVMPattern<memref::RankOp>::ConvertOpToLLVMPattern; |
| 1045 | |
| 1046 | LogicalResult |
| 1047 | matchAndRewrite(memref::RankOp op, OpAdaptor adaptor, |
| 1048 | ConversionPatternRewriter &rewriter) const override { |
| 1049 | Location loc = op.getLoc(); |
| 1050 | Type operandType = op.getMemref().getType(); |
| 1051 | if (isa<UnrankedMemRefType>(Val: operandType)) { |
| 1052 | UnrankedMemRefDescriptor desc(adaptor.getMemref()); |
| 1053 | rewriter.replaceOp(op, newValues: {desc.rank(builder&: rewriter, loc)}); |
| 1054 | return success(); |
| 1055 | } |
| 1056 | if (auto rankedMemRefType = dyn_cast<MemRefType>(Val&: operandType)) { |
| 1057 | Type indexType = getIndexType(); |
| 1058 | rewriter.replaceOp(op, |
| 1059 | newValues: {createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, |
| 1060 | value: rankedMemRefType.getRank())}); |
| 1061 | return success(); |
| 1062 | } |
| 1063 | return failure(); |
| 1064 | } |
| 1065 | }; |
| 1066 | |
| 1067 | struct MemRefCastOpLowering : public ConvertOpToLLVMPattern<memref::CastOp> { |
| 1068 | using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern; |
| 1069 | |
| 1070 | LogicalResult |
| 1071 | matchAndRewrite(memref::CastOp memRefCastOp, OpAdaptor adaptor, |
| 1072 | ConversionPatternRewriter &rewriter) const override { |
| 1073 | Type srcType = memRefCastOp.getOperand().getType(); |
| 1074 | Type dstType = memRefCastOp.getType(); |
| 1075 | |
| 1076 | // memref::CastOp reduce to bitcast in the ranked MemRef case and can be |
| 1077 | // used for type erasure. For now they must preserve underlying element type |
| 1078 | // and require source and result type to have the same rank. Therefore, |
| 1079 | // perform a sanity check that the underlying structs are the same. Once op |
| 1080 | // semantics are relaxed we can revisit. |
| 1081 | if (isa<MemRefType>(Val: srcType) && isa<MemRefType>(Val: dstType)) |
| 1082 | if (typeConverter->convertType(t: srcType) != |
| 1083 | typeConverter->convertType(t: dstType)) |
| 1084 | return failure(); |
| 1085 | |
| 1086 | // Unranked to unranked cast is disallowed |
| 1087 | if (isa<UnrankedMemRefType>(Val: srcType) && isa<UnrankedMemRefType>(Val: dstType)) |
| 1088 | return failure(); |
| 1089 | |
| 1090 | auto targetStructType = typeConverter->convertType(t: memRefCastOp.getType()); |
| 1091 | auto loc = memRefCastOp.getLoc(); |
| 1092 | |
| 1093 | // For ranked/ranked case, just keep the original descriptor. |
| 1094 | if (isa<MemRefType>(Val: srcType) && isa<MemRefType>(Val: dstType)) { |
| 1095 | rewriter.replaceOp(op: memRefCastOp, newValues: {adaptor.getSource()}); |
| 1096 | return success(); |
| 1097 | } |
| 1098 | |
| 1099 | if (isa<MemRefType>(Val: srcType) && isa<UnrankedMemRefType>(Val: dstType)) { |
| 1100 | // Casting ranked to unranked memref type |
| 1101 | // Set the rank in the destination from the memref type |
| 1102 | // Allocate space on the stack and copy the src memref descriptor |
| 1103 | // Set the ptr in the destination to the stack space |
| 1104 | auto srcMemRefType = cast<MemRefType>(Val&: srcType); |
| 1105 | int64_t rank = srcMemRefType.getRank(); |
| 1106 | // ptr = AllocaOp sizeof(MemRefDescriptor) |
| 1107 | auto ptr = getTypeConverter()->promoteOneMemRefDescriptor( |
| 1108 | loc, operand: adaptor.getSource(), builder&: rewriter); |
| 1109 | |
| 1110 | // rank = ConstantOp srcRank |
| 1111 | auto rankVal = |
| 1112 | createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: rank); |
| 1113 | // poison = PoisonOp |
| 1114 | UnrankedMemRefDescriptor memRefDesc = |
| 1115 | UnrankedMemRefDescriptor::poison(builder&: rewriter, loc, descriptorType: targetStructType); |
| 1116 | // d1 = InsertValueOp poison, rank, 0 |
| 1117 | memRefDesc.setRank(builder&: rewriter, loc, value: rankVal); |
| 1118 | // d2 = InsertValueOp d1, ptr, 1 |
| 1119 | memRefDesc.setMemRefDescPtr(builder&: rewriter, loc, value: ptr); |
| 1120 | rewriter.replaceOp(op: memRefCastOp, newValues: (Value)memRefDesc); |
| 1121 | |
| 1122 | } else if (isa<UnrankedMemRefType>(Val: srcType) && isa<MemRefType>(Val: dstType)) { |
| 1123 | // Casting from unranked type to ranked. |
| 1124 | // The operation is assumed to be doing a correct cast. If the destination |
| 1125 | // type mismatches the unranked the type, it is undefined behavior. |
| 1126 | UnrankedMemRefDescriptor memRefDesc(adaptor.getSource()); |
| 1127 | // ptr = ExtractValueOp src, 1 |
| 1128 | auto ptr = memRefDesc.memRefDescPtr(builder&: rewriter, loc); |
| 1129 | |
| 1130 | // struct = LoadOp ptr |
| 1131 | auto loadOp = LLVM::LoadOp::create(builder&: rewriter, location: loc, type: targetStructType, addr: ptr); |
| 1132 | rewriter.replaceOp(op: memRefCastOp, newValues: loadOp.getResult()); |
| 1133 | } else { |
| 1134 | llvm_unreachable("Unsupported unranked memref to unranked memref cast" ); |
| 1135 | } |
| 1136 | |
| 1137 | return success(); |
| 1138 | } |
| 1139 | }; |
| 1140 | |
| 1141 | /// Pattern to lower a `memref.copy` to llvm. |
| 1142 | /// |
| 1143 | /// For memrefs with identity layouts, the copy is lowered to the llvm |
| 1144 | /// `memcpy` intrinsic. For non-identity layouts, the copy is lowered to a call |
| 1145 | /// to the generic `MemrefCopyFn`. |
| 1146 | class MemRefCopyOpLowering : public ConvertOpToLLVMPattern<memref::CopyOp> { |
| 1147 | SymbolTableCollection *symbolTables = nullptr; |
| 1148 | |
| 1149 | public: |
| 1150 | explicit MemRefCopyOpLowering(const LLVMTypeConverter &typeConverter, |
| 1151 | SymbolTableCollection *symbolTables = nullptr, |
| 1152 | PatternBenefit benefit = 1) |
| 1153 | : ConvertOpToLLVMPattern<memref::CopyOp>(typeConverter, benefit), |
| 1154 | symbolTables(symbolTables) {} |
| 1155 | |
| 1156 | LogicalResult |
| 1157 | lowerToMemCopyIntrinsic(memref::CopyOp op, OpAdaptor adaptor, |
| 1158 | ConversionPatternRewriter &rewriter) const { |
| 1159 | auto loc = op.getLoc(); |
| 1160 | auto srcType = dyn_cast<MemRefType>(Val: op.getSource().getType()); |
| 1161 | |
| 1162 | MemRefDescriptor srcDesc(adaptor.getSource()); |
| 1163 | |
| 1164 | // Compute number of elements. |
| 1165 | Value numElements = |
| 1166 | createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: 1); |
| 1167 | for (int pos = 0; pos < srcType.getRank(); ++pos) { |
| 1168 | auto size = srcDesc.size(builder&: rewriter, loc, pos); |
| 1169 | numElements = LLVM::MulOp::create(builder&: rewriter, location: loc, lhs: numElements, rhs: size); |
| 1170 | } |
| 1171 | |
| 1172 | // Get element size. |
| 1173 | auto sizeInBytes = getSizeInBytes(loc, type: srcType.getElementType(), rewriter); |
| 1174 | // Compute total. |
| 1175 | Value totalSize = |
| 1176 | LLVM::MulOp::create(builder&: rewriter, location: loc, lhs: numElements, rhs: sizeInBytes); |
| 1177 | |
| 1178 | Type elementType = typeConverter->convertType(t: srcType.getElementType()); |
| 1179 | |
| 1180 | Value srcBasePtr = srcDesc.alignedPtr(builder&: rewriter, loc); |
| 1181 | Value srcOffset = srcDesc.offset(builder&: rewriter, loc); |
| 1182 | Value srcPtr = LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: srcBasePtr.getType(), |
| 1183 | elementType, basePtr: srcBasePtr, indices: srcOffset); |
| 1184 | MemRefDescriptor targetDesc(adaptor.getTarget()); |
| 1185 | Value targetBasePtr = targetDesc.alignedPtr(builder&: rewriter, loc); |
| 1186 | Value targetOffset = targetDesc.offset(builder&: rewriter, loc); |
| 1187 | Value targetPtr = |
| 1188 | LLVM::GEPOp::create(builder&: rewriter, location: loc, resultType: targetBasePtr.getType(), elementType, |
| 1189 | basePtr: targetBasePtr, indices: targetOffset); |
| 1190 | LLVM::MemcpyOp::create(builder&: rewriter, location: loc, dst: targetPtr, src: srcPtr, len: totalSize, |
| 1191 | /*isVolatile=*/false); |
| 1192 | rewriter.eraseOp(op); |
| 1193 | |
| 1194 | return success(); |
| 1195 | } |
| 1196 | |
| 1197 | LogicalResult |
| 1198 | lowerToMemCopyFunctionCall(memref::CopyOp op, OpAdaptor adaptor, |
| 1199 | ConversionPatternRewriter &rewriter) const { |
| 1200 | auto loc = op.getLoc(); |
| 1201 | auto srcType = cast<BaseMemRefType>(Val: op.getSource().getType()); |
| 1202 | auto targetType = cast<BaseMemRefType>(Val: op.getTarget().getType()); |
| 1203 | |
| 1204 | // First make sure we have an unranked memref descriptor representation. |
| 1205 | auto makeUnranked = [&, this](Value ranked, MemRefType type) { |
| 1206 | auto rank = LLVM::ConstantOp::create(builder&: rewriter, location: loc, type: getIndexType(), |
| 1207 | value: type.getRank()); |
| 1208 | auto *typeConverter = getTypeConverter(); |
| 1209 | auto ptr = |
| 1210 | typeConverter->promoteOneMemRefDescriptor(loc, operand: ranked, builder&: rewriter); |
| 1211 | |
| 1212 | auto unrankedType = |
| 1213 | UnrankedMemRefType::get(elementType: type.getElementType(), memorySpace: type.getMemorySpace()); |
| 1214 | return UnrankedMemRefDescriptor::pack( |
| 1215 | builder&: rewriter, loc, converter: *typeConverter, type: unrankedType, values: ValueRange{rank, ptr}); |
| 1216 | }; |
| 1217 | |
| 1218 | // Save stack position before promoting descriptors |
| 1219 | auto stackSaveOp = LLVM::StackSaveOp::create(builder&: rewriter, location: loc, res: getPtrType()); |
| 1220 | |
| 1221 | auto srcMemRefType = dyn_cast<MemRefType>(Val&: srcType); |
| 1222 | Value unrankedSource = |
| 1223 | srcMemRefType ? makeUnranked(adaptor.getSource(), srcMemRefType) |
| 1224 | : adaptor.getSource(); |
| 1225 | auto targetMemRefType = dyn_cast<MemRefType>(Val&: targetType); |
| 1226 | Value unrankedTarget = |
| 1227 | targetMemRefType ? makeUnranked(adaptor.getTarget(), targetMemRefType) |
| 1228 | : adaptor.getTarget(); |
| 1229 | |
| 1230 | // Now promote the unranked descriptors to the stack. |
| 1231 | auto one = createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: 1); |
| 1232 | auto promote = [&](Value desc) { |
| 1233 | auto ptrType = LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 1234 | auto allocated = |
| 1235 | LLVM::AllocaOp::create(builder&: rewriter, location: loc, resultType: ptrType, elementType: desc.getType(), arraySize: one); |
| 1236 | LLVM::StoreOp::create(builder&: rewriter, location: loc, value: desc, addr: allocated); |
| 1237 | return allocated; |
| 1238 | }; |
| 1239 | |
| 1240 | auto sourcePtr = promote(unrankedSource); |
| 1241 | auto targetPtr = promote(unrankedTarget); |
| 1242 | |
| 1243 | // Derive size from llvm.getelementptr which will account for any |
| 1244 | // potential alignment |
| 1245 | auto elemSize = getSizeInBytes(loc, type: srcType.getElementType(), rewriter); |
| 1246 | auto copyFn = LLVM::lookupOrCreateMemRefCopyFn( |
| 1247 | b&: rewriter, moduleOp: op->getParentOfType<ModuleOp>(), indexType: getIndexType(), |
| 1248 | unrankedDescriptorType: sourcePtr.getType(), symbolTables); |
| 1249 | if (failed(Result: copyFn)) |
| 1250 | return failure(); |
| 1251 | LLVM::CallOp::create(builder&: rewriter, location: loc, func: copyFn.value(), |
| 1252 | args: ValueRange{elemSize, sourcePtr, targetPtr}); |
| 1253 | |
| 1254 | // Restore stack used for descriptors |
| 1255 | LLVM::StackRestoreOp::create(builder&: rewriter, location: loc, ptr: stackSaveOp); |
| 1256 | |
| 1257 | rewriter.eraseOp(op); |
| 1258 | |
| 1259 | return success(); |
| 1260 | } |
| 1261 | |
| 1262 | LogicalResult |
| 1263 | matchAndRewrite(memref::CopyOp op, OpAdaptor adaptor, |
| 1264 | ConversionPatternRewriter &rewriter) const override { |
| 1265 | auto srcType = cast<BaseMemRefType>(Val: op.getSource().getType()); |
| 1266 | auto targetType = cast<BaseMemRefType>(Val: op.getTarget().getType()); |
| 1267 | |
| 1268 | auto isContiguousMemrefType = [&](BaseMemRefType type) { |
| 1269 | auto memrefType = dyn_cast<mlir::MemRefType>(Val&: type); |
| 1270 | // We can use memcpy for memrefs if they have an identity layout or are |
| 1271 | // contiguous with an arbitrary offset. Ignore empty memrefs, which is a |
| 1272 | // special case handled by memrefCopy. |
| 1273 | return memrefType && |
| 1274 | (memrefType.getLayout().isIdentity() || |
| 1275 | (memrefType.hasStaticShape() && memrefType.getNumElements() > 0 && |
| 1276 | memref::isStaticShapeAndContiguousRowMajor(type: memrefType))); |
| 1277 | }; |
| 1278 | |
| 1279 | if (isContiguousMemrefType(srcType) && isContiguousMemrefType(targetType)) |
| 1280 | return lowerToMemCopyIntrinsic(op, adaptor, rewriter); |
| 1281 | |
| 1282 | return lowerToMemCopyFunctionCall(op, adaptor, rewriter); |
| 1283 | } |
| 1284 | }; |
| 1285 | |
| 1286 | struct MemorySpaceCastOpLowering |
| 1287 | : public ConvertOpToLLVMPattern<memref::MemorySpaceCastOp> { |
| 1288 | using ConvertOpToLLVMPattern< |
| 1289 | memref::MemorySpaceCastOp>::ConvertOpToLLVMPattern; |
| 1290 | |
| 1291 | LogicalResult |
| 1292 | matchAndRewrite(memref::MemorySpaceCastOp op, OpAdaptor adaptor, |
| 1293 | ConversionPatternRewriter &rewriter) const override { |
| 1294 | Location loc = op.getLoc(); |
| 1295 | |
| 1296 | Type resultType = op.getDest().getType(); |
| 1297 | if (auto resultTypeR = dyn_cast<MemRefType>(Val&: resultType)) { |
| 1298 | auto convertedType = |
| 1299 | typeConverter->convertType<LLVM::LLVMStructType>(t: resultTypeR); |
| 1300 | if (!convertedType) |
| 1301 | return rewriter.notifyMatchFailure(arg&: op, msg: "memref type conversion failed" ); |
| 1302 | Type newPtrType = convertedType.getBody()[0]; |
| 1303 | |
| 1304 | SmallVector<Value> descVals; |
| 1305 | MemRefDescriptor::unpack(builder&: rewriter, loc, packed: adaptor.getSource(), type: resultTypeR, |
| 1306 | results&: descVals); |
| 1307 | descVals[0] = |
| 1308 | LLVM::AddrSpaceCastOp::create(builder&: rewriter, location: loc, res: newPtrType, arg: descVals[0]); |
| 1309 | descVals[1] = |
| 1310 | LLVM::AddrSpaceCastOp::create(builder&: rewriter, location: loc, res: newPtrType, arg: descVals[1]); |
| 1311 | Value result = MemRefDescriptor::pack(builder&: rewriter, loc, converter: *getTypeConverter(), |
| 1312 | type: resultTypeR, values: descVals); |
| 1313 | rewriter.replaceOp(op, newValues: result); |
| 1314 | return success(); |
| 1315 | } |
| 1316 | if (auto resultTypeU = dyn_cast<UnrankedMemRefType>(Val&: resultType)) { |
| 1317 | // Since the type converter won't be doing this for us, get the address |
| 1318 | // space. |
| 1319 | auto sourceType = cast<UnrankedMemRefType>(Val: op.getSource().getType()); |
| 1320 | FailureOr<unsigned> maybeSourceAddrSpace = |
| 1321 | getTypeConverter()->getMemRefAddressSpace(type: sourceType); |
| 1322 | if (failed(Result: maybeSourceAddrSpace)) |
| 1323 | return rewriter.notifyMatchFailure(arg&: loc, |
| 1324 | msg: "non-integer source address space" ); |
| 1325 | unsigned sourceAddrSpace = *maybeSourceAddrSpace; |
| 1326 | FailureOr<unsigned> maybeResultAddrSpace = |
| 1327 | getTypeConverter()->getMemRefAddressSpace(type: resultTypeU); |
| 1328 | if (failed(Result: maybeResultAddrSpace)) |
| 1329 | return rewriter.notifyMatchFailure(arg&: loc, |
| 1330 | msg: "non-integer result address space" ); |
| 1331 | unsigned resultAddrSpace = *maybeResultAddrSpace; |
| 1332 | |
| 1333 | UnrankedMemRefDescriptor sourceDesc(adaptor.getSource()); |
| 1334 | Value rank = sourceDesc.rank(builder&: rewriter, loc); |
| 1335 | Value sourceUnderlyingDesc = sourceDesc.memRefDescPtr(builder&: rewriter, loc); |
| 1336 | |
| 1337 | // Create and allocate storage for new memref descriptor. |
| 1338 | auto result = UnrankedMemRefDescriptor::poison( |
| 1339 | builder&: rewriter, loc, descriptorType: typeConverter->convertType(t: resultTypeU)); |
| 1340 | result.setRank(builder&: rewriter, loc, value: rank); |
| 1341 | Value resultUnderlyingSize = UnrankedMemRefDescriptor::computeSize( |
| 1342 | builder&: rewriter, loc, typeConverter: *getTypeConverter(), desc: result, addressSpace: resultAddrSpace); |
| 1343 | Value resultUnderlyingDesc = |
| 1344 | LLVM::AllocaOp::create(builder&: rewriter, location: loc, resultType: getPtrType(), |
| 1345 | elementType: rewriter.getI8Type(), arraySize: resultUnderlyingSize); |
| 1346 | result.setMemRefDescPtr(builder&: rewriter, loc, value: resultUnderlyingDesc); |
| 1347 | |
| 1348 | // Copy pointers, performing address space casts. |
| 1349 | auto sourceElemPtrType = |
| 1350 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: sourceAddrSpace); |
| 1351 | auto resultElemPtrType = |
| 1352 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: resultAddrSpace); |
| 1353 | |
| 1354 | Value allocatedPtr = sourceDesc.allocatedPtr( |
| 1355 | builder&: rewriter, loc, memRefDescPtr: sourceUnderlyingDesc, elemPtrType: sourceElemPtrType); |
| 1356 | Value alignedPtr = |
| 1357 | sourceDesc.alignedPtr(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1358 | memRefDescPtr: sourceUnderlyingDesc, elemPtrType: sourceElemPtrType); |
| 1359 | allocatedPtr = LLVM::AddrSpaceCastOp::create( |
| 1360 | builder&: rewriter, location: loc, res: resultElemPtrType, arg: allocatedPtr); |
| 1361 | alignedPtr = LLVM::AddrSpaceCastOp::create(builder&: rewriter, location: loc, |
| 1362 | res: resultElemPtrType, arg: alignedPtr); |
| 1363 | |
| 1364 | result.setAllocatedPtr(builder&: rewriter, loc, memRefDescPtr: resultUnderlyingDesc, |
| 1365 | elemPtrType: resultElemPtrType, allocatedPtr); |
| 1366 | result.setAlignedPtr(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1367 | memRefDescPtr: resultUnderlyingDesc, elemPtrType: resultElemPtrType, alignedPtr); |
| 1368 | |
| 1369 | // Copy all the index-valued operands. |
| 1370 | Value sourceIndexVals = |
| 1371 | sourceDesc.offsetBasePtr(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1372 | memRefDescPtr: sourceUnderlyingDesc, elemPtrType: sourceElemPtrType); |
| 1373 | Value resultIndexVals = |
| 1374 | result.offsetBasePtr(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1375 | memRefDescPtr: resultUnderlyingDesc, elemPtrType: resultElemPtrType); |
| 1376 | |
| 1377 | int64_t bytesToSkip = |
| 1378 | 2 * llvm::divideCeil( |
| 1379 | Numerator: getTypeConverter()->getPointerBitwidth(addressSpace: resultAddrSpace), Denominator: 8); |
| 1380 | Value bytesToSkipConst = |
| 1381 | createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: bytesToSkip); |
| 1382 | Value copySize = |
| 1383 | LLVM::SubOp::create(builder&: rewriter, location: loc, res: getIndexType(), |
| 1384 | lhs: resultUnderlyingSize, rhs: bytesToSkipConst); |
| 1385 | LLVM::MemcpyOp::create(builder&: rewriter, location: loc, dst: resultIndexVals, src: sourceIndexVals, |
| 1386 | len: copySize, /*isVolatile=*/false); |
| 1387 | |
| 1388 | rewriter.replaceOp(op, newValues: ValueRange{result}); |
| 1389 | return success(); |
| 1390 | } |
| 1391 | return rewriter.notifyMatchFailure(arg&: loc, msg: "unexpected memref type" ); |
| 1392 | } |
| 1393 | }; |
| 1394 | |
| 1395 | /// Extracts allocated, aligned pointers and offset from a ranked or unranked |
| 1396 | /// memref type. In unranked case, the fields are extracted from the underlying |
| 1397 | /// ranked descriptor. |
| 1398 | static void extractPointersAndOffset(Location loc, |
| 1399 | ConversionPatternRewriter &rewriter, |
| 1400 | const LLVMTypeConverter &typeConverter, |
| 1401 | Value originalOperand, |
| 1402 | Value convertedOperand, |
| 1403 | Value *allocatedPtr, Value *alignedPtr, |
| 1404 | Value *offset = nullptr) { |
| 1405 | Type operandType = originalOperand.getType(); |
| 1406 | if (isa<MemRefType>(Val: operandType)) { |
| 1407 | MemRefDescriptor desc(convertedOperand); |
| 1408 | *allocatedPtr = desc.allocatedPtr(builder&: rewriter, loc); |
| 1409 | *alignedPtr = desc.alignedPtr(builder&: rewriter, loc); |
| 1410 | if (offset != nullptr) |
| 1411 | *offset = desc.offset(builder&: rewriter, loc); |
| 1412 | return; |
| 1413 | } |
| 1414 | |
| 1415 | // These will all cause assert()s on unconvertible types. |
| 1416 | unsigned memorySpace = *typeConverter.getMemRefAddressSpace( |
| 1417 | type: cast<UnrankedMemRefType>(Val&: operandType)); |
| 1418 | auto elementPtrType = |
| 1419 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace: memorySpace); |
| 1420 | |
| 1421 | // Extract pointer to the underlying ranked memref descriptor and cast it to |
| 1422 | // ElemType**. |
| 1423 | UnrankedMemRefDescriptor unrankedDesc(convertedOperand); |
| 1424 | Value underlyingDescPtr = unrankedDesc.memRefDescPtr(builder&: rewriter, loc); |
| 1425 | |
| 1426 | *allocatedPtr = UnrankedMemRefDescriptor::allocatedPtr( |
| 1427 | builder&: rewriter, loc, memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType); |
| 1428 | *alignedPtr = UnrankedMemRefDescriptor::alignedPtr( |
| 1429 | builder&: rewriter, loc, typeConverter, memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType); |
| 1430 | if (offset != nullptr) { |
| 1431 | *offset = UnrankedMemRefDescriptor::offset( |
| 1432 | builder&: rewriter, loc, typeConverter, memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType); |
| 1433 | } |
| 1434 | } |
| 1435 | |
| 1436 | struct MemRefReinterpretCastOpLowering |
| 1437 | : public ConvertOpToLLVMPattern<memref::ReinterpretCastOp> { |
| 1438 | using ConvertOpToLLVMPattern< |
| 1439 | memref::ReinterpretCastOp>::ConvertOpToLLVMPattern; |
| 1440 | |
| 1441 | LogicalResult |
| 1442 | matchAndRewrite(memref::ReinterpretCastOp castOp, OpAdaptor adaptor, |
| 1443 | ConversionPatternRewriter &rewriter) const override { |
| 1444 | Type srcType = castOp.getSource().getType(); |
| 1445 | |
| 1446 | Value descriptor; |
| 1447 | if (failed(Result: convertSourceMemRefToDescriptor(rewriter, srcType, castOp, |
| 1448 | adaptor, descriptor: &descriptor))) |
| 1449 | return failure(); |
| 1450 | rewriter.replaceOp(op: castOp, newValues: {descriptor}); |
| 1451 | return success(); |
| 1452 | } |
| 1453 | |
| 1454 | private: |
| 1455 | LogicalResult convertSourceMemRefToDescriptor( |
| 1456 | ConversionPatternRewriter &rewriter, Type srcType, |
| 1457 | memref::ReinterpretCastOp castOp, |
| 1458 | memref::ReinterpretCastOp::Adaptor adaptor, Value *descriptor) const { |
| 1459 | MemRefType targetMemRefType = |
| 1460 | cast<MemRefType>(Val: castOp.getResult().getType()); |
| 1461 | auto llvmTargetDescriptorTy = |
| 1462 | typeConverter->convertType<LLVM::LLVMStructType>(t: targetMemRefType); |
| 1463 | if (!llvmTargetDescriptorTy) |
| 1464 | return failure(); |
| 1465 | |
| 1466 | // Create descriptor. |
| 1467 | Location loc = castOp.getLoc(); |
| 1468 | auto desc = MemRefDescriptor::poison(builder&: rewriter, loc, descriptorType: llvmTargetDescriptorTy); |
| 1469 | |
| 1470 | // Set allocated and aligned pointers. |
| 1471 | Value allocatedPtr, alignedPtr; |
| 1472 | extractPointersAndOffset(loc, rewriter, typeConverter: *getTypeConverter(), |
| 1473 | originalOperand: castOp.getSource(), convertedOperand: adaptor.getSource(), |
| 1474 | allocatedPtr: &allocatedPtr, alignedPtr: &alignedPtr); |
| 1475 | desc.setAllocatedPtr(builder&: rewriter, loc, ptr: allocatedPtr); |
| 1476 | desc.setAlignedPtr(builder&: rewriter, loc, ptr: alignedPtr); |
| 1477 | |
| 1478 | // Set offset. |
| 1479 | if (castOp.isDynamicOffset(idx: 0)) |
| 1480 | desc.setOffset(builder&: rewriter, loc, offset: adaptor.getOffsets()[0]); |
| 1481 | else |
| 1482 | desc.setConstantOffset(builder&: rewriter, loc, offset: castOp.getStaticOffset(idx: 0)); |
| 1483 | |
| 1484 | // Set sizes and strides. |
| 1485 | unsigned dynSizeId = 0; |
| 1486 | unsigned dynStrideId = 0; |
| 1487 | for (unsigned i = 0, e = targetMemRefType.getRank(); i < e; ++i) { |
| 1488 | if (castOp.isDynamicSize(idx: i)) |
| 1489 | desc.setSize(builder&: rewriter, loc, pos: i, size: adaptor.getSizes()[dynSizeId++]); |
| 1490 | else |
| 1491 | desc.setConstantSize(builder&: rewriter, loc, pos: i, size: castOp.getStaticSize(idx: i)); |
| 1492 | |
| 1493 | if (castOp.isDynamicStride(idx: i)) |
| 1494 | desc.setStride(builder&: rewriter, loc, pos: i, stride: adaptor.getStrides()[dynStrideId++]); |
| 1495 | else |
| 1496 | desc.setConstantStride(builder&: rewriter, loc, pos: i, stride: castOp.getStaticStride(idx: i)); |
| 1497 | } |
| 1498 | *descriptor = desc; |
| 1499 | return success(); |
| 1500 | } |
| 1501 | }; |
| 1502 | |
| 1503 | struct MemRefReshapeOpLowering |
| 1504 | : public ConvertOpToLLVMPattern<memref::ReshapeOp> { |
| 1505 | using ConvertOpToLLVMPattern<memref::ReshapeOp>::ConvertOpToLLVMPattern; |
| 1506 | |
| 1507 | LogicalResult |
| 1508 | matchAndRewrite(memref::ReshapeOp reshapeOp, OpAdaptor adaptor, |
| 1509 | ConversionPatternRewriter &rewriter) const override { |
| 1510 | Type srcType = reshapeOp.getSource().getType(); |
| 1511 | |
| 1512 | Value descriptor; |
| 1513 | if (failed(Result: convertSourceMemRefToDescriptor(rewriter, srcType, reshapeOp, |
| 1514 | adaptor, descriptor: &descriptor))) |
| 1515 | return failure(); |
| 1516 | rewriter.replaceOp(op: reshapeOp, newValues: {descriptor}); |
| 1517 | return success(); |
| 1518 | } |
| 1519 | |
| 1520 | private: |
| 1521 | LogicalResult |
| 1522 | convertSourceMemRefToDescriptor(ConversionPatternRewriter &rewriter, |
| 1523 | Type srcType, memref::ReshapeOp reshapeOp, |
| 1524 | memref::ReshapeOp::Adaptor adaptor, |
| 1525 | Value *descriptor) const { |
| 1526 | auto shapeMemRefType = cast<MemRefType>(Val: reshapeOp.getShape().getType()); |
| 1527 | if (shapeMemRefType.hasStaticShape()) { |
| 1528 | MemRefType targetMemRefType = |
| 1529 | cast<MemRefType>(Val: reshapeOp.getResult().getType()); |
| 1530 | auto llvmTargetDescriptorTy = |
| 1531 | typeConverter->convertType<LLVM::LLVMStructType>(t: targetMemRefType); |
| 1532 | if (!llvmTargetDescriptorTy) |
| 1533 | return failure(); |
| 1534 | |
| 1535 | // Create descriptor. |
| 1536 | Location loc = reshapeOp.getLoc(); |
| 1537 | auto desc = |
| 1538 | MemRefDescriptor::poison(builder&: rewriter, loc, descriptorType: llvmTargetDescriptorTy); |
| 1539 | |
| 1540 | // Set allocated and aligned pointers. |
| 1541 | Value allocatedPtr, alignedPtr; |
| 1542 | extractPointersAndOffset(loc, rewriter, typeConverter: *getTypeConverter(), |
| 1543 | originalOperand: reshapeOp.getSource(), convertedOperand: adaptor.getSource(), |
| 1544 | allocatedPtr: &allocatedPtr, alignedPtr: &alignedPtr); |
| 1545 | desc.setAllocatedPtr(builder&: rewriter, loc, ptr: allocatedPtr); |
| 1546 | desc.setAlignedPtr(builder&: rewriter, loc, ptr: alignedPtr); |
| 1547 | |
| 1548 | // Extract the offset and strides from the type. |
| 1549 | int64_t offset; |
| 1550 | SmallVector<int64_t> strides; |
| 1551 | if (failed(Result: targetMemRefType.getStridesAndOffset(strides, offset))) |
| 1552 | return rewriter.notifyMatchFailure( |
| 1553 | arg&: reshapeOp, msg: "failed to get stride and offset exprs" ); |
| 1554 | |
| 1555 | if (!isStaticStrideOrOffset(strideOrOffset: offset)) |
| 1556 | return rewriter.notifyMatchFailure(arg&: reshapeOp, |
| 1557 | msg: "dynamic offset is unsupported" ); |
| 1558 | |
| 1559 | desc.setConstantOffset(builder&: rewriter, loc, offset); |
| 1560 | |
| 1561 | assert(targetMemRefType.getLayout().isIdentity() && |
| 1562 | "Identity layout map is a precondition of a valid reshape op" ); |
| 1563 | |
| 1564 | Type indexType = getIndexType(); |
| 1565 | Value stride = nullptr; |
| 1566 | int64_t targetRank = targetMemRefType.getRank(); |
| 1567 | for (auto i : llvm::reverse(C: llvm::seq<int64_t>(Begin: 0, End: targetRank))) { |
| 1568 | if (ShapedType::isStatic(dValue: strides[i])) { |
| 1569 | // If the stride for this dimension is dynamic, then use the product |
| 1570 | // of the sizes of the inner dimensions. |
| 1571 | stride = |
| 1572 | createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: strides[i]); |
| 1573 | } else if (!stride) { |
| 1574 | // `stride` is null only in the first iteration of the loop. However, |
| 1575 | // since the target memref has an identity layout, we can safely set |
| 1576 | // the innermost stride to 1. |
| 1577 | stride = createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: 1); |
| 1578 | } |
| 1579 | |
| 1580 | Value dimSize; |
| 1581 | // If the size of this dimension is dynamic, then load it at runtime |
| 1582 | // from the shape operand. |
| 1583 | if (!targetMemRefType.isDynamicDim(idx: i)) { |
| 1584 | dimSize = createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, |
| 1585 | value: targetMemRefType.getDimSize(idx: i)); |
| 1586 | } else { |
| 1587 | Value shapeOp = reshapeOp.getShape(); |
| 1588 | Value index = createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: i); |
| 1589 | dimSize = memref::LoadOp::create(builder&: rewriter, location: loc, memref: shapeOp, indices: index); |
| 1590 | Type indexType = getIndexType(); |
| 1591 | if (dimSize.getType() != indexType) |
| 1592 | dimSize = typeConverter->materializeTargetConversion( |
| 1593 | builder&: rewriter, loc, resultType: indexType, inputs: dimSize); |
| 1594 | assert(dimSize && "Invalid memref element type" ); |
| 1595 | } |
| 1596 | |
| 1597 | desc.setSize(builder&: rewriter, loc, pos: i, size: dimSize); |
| 1598 | desc.setStride(builder&: rewriter, loc, pos: i, stride); |
| 1599 | |
| 1600 | // Prepare the stride value for the next dimension. |
| 1601 | stride = LLVM::MulOp::create(builder&: rewriter, location: loc, lhs: stride, rhs: dimSize); |
| 1602 | } |
| 1603 | |
| 1604 | *descriptor = desc; |
| 1605 | return success(); |
| 1606 | } |
| 1607 | |
| 1608 | // The shape is a rank-1 tensor with unknown length. |
| 1609 | Location loc = reshapeOp.getLoc(); |
| 1610 | MemRefDescriptor shapeDesc(adaptor.getShape()); |
| 1611 | Value resultRank = shapeDesc.size(builder&: rewriter, loc, pos: 0); |
| 1612 | |
| 1613 | // Extract address space and element type. |
| 1614 | auto targetType = cast<UnrankedMemRefType>(Val: reshapeOp.getResult().getType()); |
| 1615 | unsigned addressSpace = |
| 1616 | *getTypeConverter()->getMemRefAddressSpace(type: targetType); |
| 1617 | |
| 1618 | // Create the unranked memref descriptor that holds the ranked one. The |
| 1619 | // inner descriptor is allocated on stack. |
| 1620 | auto targetDesc = UnrankedMemRefDescriptor::poison( |
| 1621 | builder&: rewriter, loc, descriptorType: typeConverter->convertType(t: targetType)); |
| 1622 | targetDesc.setRank(builder&: rewriter, loc, value: resultRank); |
| 1623 | Value allocationSize = UnrankedMemRefDescriptor::computeSize( |
| 1624 | builder&: rewriter, loc, typeConverter: *getTypeConverter(), desc: targetDesc, addressSpace); |
| 1625 | Value underlyingDescPtr = LLVM::AllocaOp::create( |
| 1626 | builder&: rewriter, location: loc, resultType: getPtrType(), elementType: IntegerType::get(context: getContext(), width: 8), |
| 1627 | arraySize: allocationSize); |
| 1628 | targetDesc.setMemRefDescPtr(builder&: rewriter, loc, value: underlyingDescPtr); |
| 1629 | |
| 1630 | // Extract pointers and offset from the source memref. |
| 1631 | Value allocatedPtr, alignedPtr, offset; |
| 1632 | extractPointersAndOffset(loc, rewriter, typeConverter: *getTypeConverter(), |
| 1633 | originalOperand: reshapeOp.getSource(), convertedOperand: adaptor.getSource(), |
| 1634 | allocatedPtr: &allocatedPtr, alignedPtr: &alignedPtr, offset: &offset); |
| 1635 | |
| 1636 | // Set pointers and offset. |
| 1637 | auto elementPtrType = |
| 1638 | LLVM::LLVMPointerType::get(context: rewriter.getContext(), addressSpace); |
| 1639 | |
| 1640 | UnrankedMemRefDescriptor::setAllocatedPtr(builder&: rewriter, loc, memRefDescPtr: underlyingDescPtr, |
| 1641 | elemPtrType: elementPtrType, allocatedPtr); |
| 1642 | UnrankedMemRefDescriptor::setAlignedPtr(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1643 | memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType, |
| 1644 | alignedPtr); |
| 1645 | UnrankedMemRefDescriptor::setOffset(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1646 | memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType, |
| 1647 | offset); |
| 1648 | |
| 1649 | // Use the offset pointer as base for further addressing. Copy over the new |
| 1650 | // shape and compute strides. For this, we create a loop from rank-1 to 0. |
| 1651 | Value targetSizesBase = UnrankedMemRefDescriptor::sizeBasePtr( |
| 1652 | builder&: rewriter, loc, typeConverter: *getTypeConverter(), memRefDescPtr: underlyingDescPtr, elemPtrType: elementPtrType); |
| 1653 | Value targetStridesBase = UnrankedMemRefDescriptor::strideBasePtr( |
| 1654 | builder&: rewriter, loc, typeConverter: *getTypeConverter(), sizeBasePtr: targetSizesBase, rank: resultRank); |
| 1655 | Value shapeOperandPtr = shapeDesc.alignedPtr(builder&: rewriter, loc); |
| 1656 | Value oneIndex = createIndexAttrConstant(builder&: rewriter, loc, resultType: getIndexType(), value: 1); |
| 1657 | Value resultRankMinusOne = |
| 1658 | LLVM::SubOp::create(builder&: rewriter, location: loc, lhs: resultRank, rhs: oneIndex); |
| 1659 | |
| 1660 | Block *initBlock = rewriter.getInsertionBlock(); |
| 1661 | Type indexType = getTypeConverter()->getIndexType(); |
| 1662 | Block::iterator remainingOpsIt = std::next(x: rewriter.getInsertionPoint()); |
| 1663 | |
| 1664 | Block *condBlock = rewriter.createBlock(parent: initBlock->getParent(), insertPt: {}, |
| 1665 | argTypes: {indexType, indexType}, locs: {loc, loc}); |
| 1666 | |
| 1667 | // Move the remaining initBlock ops to condBlock. |
| 1668 | Block *remainingBlock = rewriter.splitBlock(block: initBlock, before: remainingOpsIt); |
| 1669 | rewriter.mergeBlocks(source: remainingBlock, dest: condBlock, argValues: ValueRange()); |
| 1670 | |
| 1671 | rewriter.setInsertionPointToEnd(initBlock); |
| 1672 | LLVM::BrOp::create(builder&: rewriter, location: loc, |
| 1673 | operands: ValueRange({resultRankMinusOne, oneIndex}), dest: condBlock); |
| 1674 | rewriter.setInsertionPointToStart(condBlock); |
| 1675 | Value indexArg = condBlock->getArgument(i: 0); |
| 1676 | Value strideArg = condBlock->getArgument(i: 1); |
| 1677 | |
| 1678 | Value zeroIndex = createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: 0); |
| 1679 | Value pred = LLVM::ICmpOp::create( |
| 1680 | builder&: rewriter, location: loc, res: IntegerType::get(context: rewriter.getContext(), width: 1), |
| 1681 | predicate: LLVM::ICmpPredicate::sge, lhs: indexArg, rhs: zeroIndex); |
| 1682 | |
| 1683 | Block *bodyBlock = |
| 1684 | rewriter.splitBlock(block: condBlock, before: rewriter.getInsertionPoint()); |
| 1685 | rewriter.setInsertionPointToStart(bodyBlock); |
| 1686 | |
| 1687 | // Copy size from shape to descriptor. |
| 1688 | auto llvmIndexPtrType = LLVM::LLVMPointerType::get(context: rewriter.getContext()); |
| 1689 | Value sizeLoadGep = LLVM::GEPOp::create( |
| 1690 | builder&: rewriter, location: loc, resultType: llvmIndexPtrType, |
| 1691 | elementType: typeConverter->convertType(t: shapeMemRefType.getElementType()), |
| 1692 | basePtr: shapeOperandPtr, indices: indexArg); |
| 1693 | Value size = LLVM::LoadOp::create(builder&: rewriter, location: loc, type: indexType, addr: sizeLoadGep); |
| 1694 | UnrankedMemRefDescriptor::setSize(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1695 | sizeBasePtr: targetSizesBase, index: indexArg, size); |
| 1696 | |
| 1697 | // Write stride value and compute next one. |
| 1698 | UnrankedMemRefDescriptor::setStride(builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 1699 | strideBasePtr: targetStridesBase, index: indexArg, stride: strideArg); |
| 1700 | Value nextStride = LLVM::MulOp::create(builder&: rewriter, location: loc, lhs: strideArg, rhs: size); |
| 1701 | |
| 1702 | // Decrement loop counter and branch back. |
| 1703 | Value decrement = LLVM::SubOp::create(builder&: rewriter, location: loc, lhs: indexArg, rhs: oneIndex); |
| 1704 | LLVM::BrOp::create(builder&: rewriter, location: loc, operands: ValueRange({decrement, nextStride}), |
| 1705 | dest: condBlock); |
| 1706 | |
| 1707 | Block *remainder = |
| 1708 | rewriter.splitBlock(block: bodyBlock, before: rewriter.getInsertionPoint()); |
| 1709 | |
| 1710 | // Hook up the cond exit to the remainder. |
| 1711 | rewriter.setInsertionPointToEnd(condBlock); |
| 1712 | LLVM::CondBrOp::create(builder&: rewriter, location: loc, condition: pred, trueDest: bodyBlock, trueOperands: ValueRange(), |
| 1713 | falseDest: remainder, falseOperands: ValueRange()); |
| 1714 | |
| 1715 | // Reset position to beginning of new remainder block. |
| 1716 | rewriter.setInsertionPointToStart(remainder); |
| 1717 | |
| 1718 | *descriptor = targetDesc; |
| 1719 | return success(); |
| 1720 | } |
| 1721 | }; |
| 1722 | |
| 1723 | /// RessociatingReshapeOp must be expanded before we reach this stage. |
| 1724 | /// Report that information. |
| 1725 | template <typename ReshapeOp> |
| 1726 | class ReassociatingReshapeOpConversion |
| 1727 | : public ConvertOpToLLVMPattern<ReshapeOp> { |
| 1728 | public: |
| 1729 | using ConvertOpToLLVMPattern<ReshapeOp>::ConvertOpToLLVMPattern; |
| 1730 | using ReshapeOpAdaptor = typename ReshapeOp::Adaptor; |
| 1731 | |
| 1732 | LogicalResult |
| 1733 | matchAndRewrite(ReshapeOp reshapeOp, typename ReshapeOp::Adaptor adaptor, |
| 1734 | ConversionPatternRewriter &rewriter) const override { |
| 1735 | return rewriter.notifyMatchFailure( |
| 1736 | reshapeOp, |
| 1737 | "reassociation operations should have been expanded beforehand" ); |
| 1738 | } |
| 1739 | }; |
| 1740 | |
| 1741 | /// Subviews must be expanded before we reach this stage. |
| 1742 | /// Report that information. |
| 1743 | struct SubViewOpLowering : public ConvertOpToLLVMPattern<memref::SubViewOp> { |
| 1744 | using ConvertOpToLLVMPattern<memref::SubViewOp>::ConvertOpToLLVMPattern; |
| 1745 | |
| 1746 | LogicalResult |
| 1747 | matchAndRewrite(memref::SubViewOp subViewOp, OpAdaptor adaptor, |
| 1748 | ConversionPatternRewriter &rewriter) const override { |
| 1749 | return rewriter.notifyMatchFailure( |
| 1750 | arg&: subViewOp, msg: "subview operations should have been expanded beforehand" ); |
| 1751 | } |
| 1752 | }; |
| 1753 | |
| 1754 | /// Conversion pattern that transforms a transpose op into: |
| 1755 | /// 1. A function entry `alloca` operation to allocate a ViewDescriptor. |
| 1756 | /// 2. A load of the ViewDescriptor from the pointer allocated in 1. |
| 1757 | /// 3. Updates to the ViewDescriptor to introduce the data ptr, offset, size |
| 1758 | /// and stride. Size and stride are permutations of the original values. |
| 1759 | /// 4. A store of the resulting ViewDescriptor to the alloca'ed pointer. |
| 1760 | /// The transpose op is replaced by the alloca'ed pointer. |
| 1761 | class TransposeOpLowering : public ConvertOpToLLVMPattern<memref::TransposeOp> { |
| 1762 | public: |
| 1763 | using ConvertOpToLLVMPattern<memref::TransposeOp>::ConvertOpToLLVMPattern; |
| 1764 | |
| 1765 | LogicalResult |
| 1766 | matchAndRewrite(memref::TransposeOp transposeOp, OpAdaptor adaptor, |
| 1767 | ConversionPatternRewriter &rewriter) const override { |
| 1768 | auto loc = transposeOp.getLoc(); |
| 1769 | MemRefDescriptor viewMemRef(adaptor.getIn()); |
| 1770 | |
| 1771 | // No permutation, early exit. |
| 1772 | if (transposeOp.getPermutation().isIdentity()) |
| 1773 | return rewriter.replaceOp(op: transposeOp, newValues: {viewMemRef}), success(); |
| 1774 | |
| 1775 | auto targetMemRef = MemRefDescriptor::poison( |
| 1776 | builder&: rewriter, loc, |
| 1777 | descriptorType: typeConverter->convertType(t: transposeOp.getIn().getType())); |
| 1778 | |
| 1779 | // Copy the base and aligned pointers from the old descriptor to the new |
| 1780 | // one. |
| 1781 | targetMemRef.setAllocatedPtr(builder&: rewriter, loc, |
| 1782 | ptr: viewMemRef.allocatedPtr(builder&: rewriter, loc)); |
| 1783 | targetMemRef.setAlignedPtr(builder&: rewriter, loc, |
| 1784 | ptr: viewMemRef.alignedPtr(builder&: rewriter, loc)); |
| 1785 | |
| 1786 | // Copy the offset pointer from the old descriptor to the new one. |
| 1787 | targetMemRef.setOffset(builder&: rewriter, loc, offset: viewMemRef.offset(builder&: rewriter, loc)); |
| 1788 | |
| 1789 | // Iterate over the dimensions and apply size/stride permutation: |
| 1790 | // When enumerating the results of the permutation map, the enumeration |
| 1791 | // index is the index into the target dimensions and the DimExpr points to |
| 1792 | // the dimension of the source memref. |
| 1793 | for (const auto &en : |
| 1794 | llvm::enumerate(First: transposeOp.getPermutation().getResults())) { |
| 1795 | int targetPos = en.index(); |
| 1796 | int sourcePos = cast<AffineDimExpr>(Val: en.value()).getPosition(); |
| 1797 | targetMemRef.setSize(builder&: rewriter, loc, pos: targetPos, |
| 1798 | size: viewMemRef.size(builder&: rewriter, loc, pos: sourcePos)); |
| 1799 | targetMemRef.setStride(builder&: rewriter, loc, pos: targetPos, |
| 1800 | stride: viewMemRef.stride(builder&: rewriter, loc, pos: sourcePos)); |
| 1801 | } |
| 1802 | |
| 1803 | rewriter.replaceOp(op: transposeOp, newValues: {targetMemRef}); |
| 1804 | return success(); |
| 1805 | } |
| 1806 | }; |
| 1807 | |
| 1808 | /// Conversion pattern that transforms an op into: |
| 1809 | /// 1. An `llvm.mlir.undef` operation to create a memref descriptor |
| 1810 | /// 2. Updates to the descriptor to introduce the data ptr, offset, size |
| 1811 | /// and stride. |
| 1812 | /// The view op is replaced by the descriptor. |
| 1813 | struct ViewOpLowering : public ConvertOpToLLVMPattern<memref::ViewOp> { |
| 1814 | using ConvertOpToLLVMPattern<memref::ViewOp>::ConvertOpToLLVMPattern; |
| 1815 | |
| 1816 | // Build and return the value for the idx^th shape dimension, either by |
| 1817 | // returning the constant shape dimension or counting the proper dynamic size. |
| 1818 | Value getSize(ConversionPatternRewriter &rewriter, Location loc, |
| 1819 | ArrayRef<int64_t> shape, ValueRange dynamicSizes, unsigned idx, |
| 1820 | Type indexType) const { |
| 1821 | assert(idx < shape.size()); |
| 1822 | if (ShapedType::isStatic(dValue: shape[idx])) |
| 1823 | return createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: shape[idx]); |
| 1824 | // Count the number of dynamic dims in range [0, idx] |
| 1825 | unsigned nDynamic = |
| 1826 | llvm::count_if(Range: shape.take_front(N: idx), P: ShapedType::isDynamic); |
| 1827 | return dynamicSizes[nDynamic]; |
| 1828 | } |
| 1829 | |
| 1830 | // Build and return the idx^th stride, either by returning the constant stride |
| 1831 | // or by computing the dynamic stride from the current `runningStride` and |
| 1832 | // `nextSize`. The caller should keep a running stride and update it with the |
| 1833 | // result returned by this function. |
| 1834 | Value getStride(ConversionPatternRewriter &rewriter, Location loc, |
| 1835 | ArrayRef<int64_t> strides, Value nextSize, |
| 1836 | Value runningStride, unsigned idx, Type indexType) const { |
| 1837 | assert(idx < strides.size()); |
| 1838 | if (ShapedType::isStatic(dValue: strides[idx])) |
| 1839 | return createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: strides[idx]); |
| 1840 | if (nextSize) |
| 1841 | return runningStride |
| 1842 | ? LLVM::MulOp::create(builder&: rewriter, location: loc, lhs: runningStride, rhs: nextSize) |
| 1843 | : nextSize; |
| 1844 | assert(!runningStride); |
| 1845 | return createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: 1); |
| 1846 | } |
| 1847 | |
| 1848 | LogicalResult |
| 1849 | matchAndRewrite(memref::ViewOp viewOp, OpAdaptor adaptor, |
| 1850 | ConversionPatternRewriter &rewriter) const override { |
| 1851 | auto loc = viewOp.getLoc(); |
| 1852 | |
| 1853 | auto viewMemRefType = viewOp.getType(); |
| 1854 | auto targetElementTy = |
| 1855 | typeConverter->convertType(t: viewMemRefType.getElementType()); |
| 1856 | auto targetDescTy = typeConverter->convertType(t: viewMemRefType); |
| 1857 | if (!targetDescTy || !targetElementTy || |
| 1858 | !LLVM::isCompatibleType(type: targetElementTy) || |
| 1859 | !LLVM::isCompatibleType(type: targetDescTy)) |
| 1860 | return viewOp.emitWarning(message: "Target descriptor type not converted to LLVM" ), |
| 1861 | failure(); |
| 1862 | |
| 1863 | int64_t offset; |
| 1864 | SmallVector<int64_t, 4> strides; |
| 1865 | auto successStrides = viewMemRefType.getStridesAndOffset(strides, offset); |
| 1866 | if (failed(Result: successStrides)) |
| 1867 | return viewOp.emitWarning(message: "cannot cast to non-strided shape" ), failure(); |
| 1868 | assert(offset == 0 && "expected offset to be 0" ); |
| 1869 | |
| 1870 | // Target memref must be contiguous in memory (innermost stride is 1), or |
| 1871 | // empty (special case when at least one of the memref dimensions is 0). |
| 1872 | if (!strides.empty() && (strides.back() != 1 && strides.back() != 0)) |
| 1873 | return viewOp.emitWarning(message: "cannot cast to non-contiguous shape" ), |
| 1874 | failure(); |
| 1875 | |
| 1876 | // Create the descriptor. |
| 1877 | MemRefDescriptor sourceMemRef(adaptor.getSource()); |
| 1878 | auto targetMemRef = MemRefDescriptor::poison(builder&: rewriter, loc, descriptorType: targetDescTy); |
| 1879 | |
| 1880 | // Field 1: Copy the allocated pointer, used for malloc/free. |
| 1881 | Value allocatedPtr = sourceMemRef.allocatedPtr(builder&: rewriter, loc); |
| 1882 | auto srcMemRefType = cast<MemRefType>(Val: viewOp.getSource().getType()); |
| 1883 | targetMemRef.setAllocatedPtr(builder&: rewriter, loc, ptr: allocatedPtr); |
| 1884 | |
| 1885 | // Field 2: Copy the actual aligned pointer to payload. |
| 1886 | Value alignedPtr = sourceMemRef.alignedPtr(builder&: rewriter, loc); |
| 1887 | alignedPtr = LLVM::GEPOp::create( |
| 1888 | builder&: rewriter, location: loc, resultType: alignedPtr.getType(), |
| 1889 | elementType: typeConverter->convertType(t: srcMemRefType.getElementType()), basePtr: alignedPtr, |
| 1890 | indices: adaptor.getByteShift()); |
| 1891 | |
| 1892 | targetMemRef.setAlignedPtr(builder&: rewriter, loc, ptr: alignedPtr); |
| 1893 | |
| 1894 | Type indexType = getIndexType(); |
| 1895 | // Field 3: The offset in the resulting type must be 0. This is |
| 1896 | // because of the type change: an offset on srcType* may not be |
| 1897 | // expressible as an offset on dstType*. |
| 1898 | targetMemRef.setOffset( |
| 1899 | builder&: rewriter, loc, |
| 1900 | offset: createIndexAttrConstant(builder&: rewriter, loc, resultType: indexType, value: offset)); |
| 1901 | |
| 1902 | // Early exit for 0-D corner case. |
| 1903 | if (viewMemRefType.getRank() == 0) |
| 1904 | return rewriter.replaceOp(op: viewOp, newValues: {targetMemRef}), success(); |
| 1905 | |
| 1906 | // Fields 4 and 5: Update sizes and strides. |
| 1907 | Value stride = nullptr, nextSize = nullptr; |
| 1908 | for (int i = viewMemRefType.getRank() - 1; i >= 0; --i) { |
| 1909 | // Update size. |
| 1910 | Value size = getSize(rewriter, loc, shape: viewMemRefType.getShape(), |
| 1911 | dynamicSizes: adaptor.getSizes(), idx: i, indexType); |
| 1912 | targetMemRef.setSize(builder&: rewriter, loc, pos: i, size); |
| 1913 | // Update stride. |
| 1914 | stride = |
| 1915 | getStride(rewriter, loc, strides, nextSize, runningStride: stride, idx: i, indexType); |
| 1916 | targetMemRef.setStride(builder&: rewriter, loc, pos: i, stride); |
| 1917 | nextSize = size; |
| 1918 | } |
| 1919 | |
| 1920 | rewriter.replaceOp(op: viewOp, newValues: {targetMemRef}); |
| 1921 | return success(); |
| 1922 | } |
| 1923 | }; |
| 1924 | |
| 1925 | //===----------------------------------------------------------------------===// |
| 1926 | // AtomicRMWOpLowering |
| 1927 | //===----------------------------------------------------------------------===// |
| 1928 | |
| 1929 | /// Try to match the kind of a memref.atomic_rmw to determine whether to use a |
| 1930 | /// lowering to llvm.atomicrmw or fallback to llvm.cmpxchg. |
| 1931 | static std::optional<LLVM::AtomicBinOp> |
| 1932 | matchSimpleAtomicOp(memref::AtomicRMWOp atomicOp) { |
| 1933 | switch (atomicOp.getKind()) { |
| 1934 | case arith::AtomicRMWKind::addf: |
| 1935 | return LLVM::AtomicBinOp::fadd; |
| 1936 | case arith::AtomicRMWKind::addi: |
| 1937 | return LLVM::AtomicBinOp::add; |
| 1938 | case arith::AtomicRMWKind::assign: |
| 1939 | return LLVM::AtomicBinOp::xchg; |
| 1940 | case arith::AtomicRMWKind::maximumf: |
| 1941 | // TODO: remove this by end of 2025. |
| 1942 | LDBG() << "the lowering of memref.atomicrmw maximumf changed " |
| 1943 | "from fmax to fmaximum, expect more NaNs" ; |
| 1944 | return LLVM::AtomicBinOp::fmaximum; |
| 1945 | case arith::AtomicRMWKind::maxnumf: |
| 1946 | return LLVM::AtomicBinOp::fmax; |
| 1947 | case arith::AtomicRMWKind::maxs: |
| 1948 | return LLVM::AtomicBinOp::max; |
| 1949 | case arith::AtomicRMWKind::maxu: |
| 1950 | return LLVM::AtomicBinOp::umax; |
| 1951 | case arith::AtomicRMWKind::minimumf: |
| 1952 | // TODO: remove this by end of 2025. |
| 1953 | LDBG() << "the lowering of memref.atomicrmw minimum changed " |
| 1954 | "from fmin to fminimum, expect more NaNs" ; |
| 1955 | return LLVM::AtomicBinOp::fminimum; |
| 1956 | case arith::AtomicRMWKind::minnumf: |
| 1957 | return LLVM::AtomicBinOp::fmin; |
| 1958 | case arith::AtomicRMWKind::mins: |
| 1959 | return LLVM::AtomicBinOp::min; |
| 1960 | case arith::AtomicRMWKind::minu: |
| 1961 | return LLVM::AtomicBinOp::umin; |
| 1962 | case arith::AtomicRMWKind::ori: |
| 1963 | return LLVM::AtomicBinOp::_or; |
| 1964 | case arith::AtomicRMWKind::xori: |
| 1965 | return LLVM::AtomicBinOp::_xor; |
| 1966 | case arith::AtomicRMWKind::andi: |
| 1967 | return LLVM::AtomicBinOp::_and; |
| 1968 | default: |
| 1969 | return std::nullopt; |
| 1970 | } |
| 1971 | llvm_unreachable("Invalid AtomicRMWKind" ); |
| 1972 | } |
| 1973 | |
| 1974 | struct AtomicRMWOpLowering : public LoadStoreOpLowering<memref::AtomicRMWOp> { |
| 1975 | using Base::Base; |
| 1976 | |
| 1977 | LogicalResult |
| 1978 | matchAndRewrite(memref::AtomicRMWOp atomicOp, OpAdaptor adaptor, |
| 1979 | ConversionPatternRewriter &rewriter) const override { |
| 1980 | auto maybeKind = matchSimpleAtomicOp(atomicOp); |
| 1981 | if (!maybeKind) |
| 1982 | return failure(); |
| 1983 | auto memRefType = atomicOp.getMemRefType(); |
| 1984 | SmallVector<int64_t> strides; |
| 1985 | int64_t offset; |
| 1986 | if (failed(Result: memRefType.getStridesAndOffset(strides, offset))) |
| 1987 | return failure(); |
| 1988 | auto dataPtr = |
| 1989 | getStridedElementPtr(rewriter, loc: atomicOp.getLoc(), type: memRefType, |
| 1990 | memRefDesc: adaptor.getMemref(), indices: adaptor.getIndices()); |
| 1991 | rewriter.replaceOpWithNewOp<LLVM::AtomicRMWOp>( |
| 1992 | op: atomicOp, args&: *maybeKind, args&: dataPtr, args: adaptor.getValue(), |
| 1993 | args: LLVM::AtomicOrdering::acq_rel); |
| 1994 | return success(); |
| 1995 | } |
| 1996 | }; |
| 1997 | |
| 1998 | /// Unpack the pointer returned by a memref.extract_aligned_pointer_as_index. |
| 1999 | class |
| 2000 | : public ConvertOpToLLVMPattern<memref::ExtractAlignedPointerAsIndexOp> { |
| 2001 | public: |
| 2002 | using ConvertOpToLLVMPattern< |
| 2003 | memref::ExtractAlignedPointerAsIndexOp>::ConvertOpToLLVMPattern; |
| 2004 | |
| 2005 | LogicalResult |
| 2006 | matchAndRewrite(memref::ExtractAlignedPointerAsIndexOp , |
| 2007 | OpAdaptor adaptor, |
| 2008 | ConversionPatternRewriter &rewriter) const override { |
| 2009 | BaseMemRefType sourceTy = extractOp.getSource().getType(); |
| 2010 | |
| 2011 | Value alignedPtr; |
| 2012 | if (sourceTy.hasRank()) { |
| 2013 | MemRefDescriptor desc(adaptor.getSource()); |
| 2014 | alignedPtr = desc.alignedPtr(builder&: rewriter, loc: extractOp->getLoc()); |
| 2015 | } else { |
| 2016 | auto elementPtrTy = LLVM::LLVMPointerType::get( |
| 2017 | context: rewriter.getContext(), addressSpace: sourceTy.getMemorySpaceAsInt()); |
| 2018 | |
| 2019 | UnrankedMemRefDescriptor desc(adaptor.getSource()); |
| 2020 | Value descPtr = desc.memRefDescPtr(builder&: rewriter, loc: extractOp->getLoc()); |
| 2021 | |
| 2022 | alignedPtr = UnrankedMemRefDescriptor::alignedPtr( |
| 2023 | builder&: rewriter, loc: extractOp->getLoc(), typeConverter: *getTypeConverter(), memRefDescPtr: descPtr, |
| 2024 | elemPtrType: elementPtrTy); |
| 2025 | } |
| 2026 | |
| 2027 | rewriter.replaceOpWithNewOp<LLVM::PtrToIntOp>( |
| 2028 | op: extractOp, args: getTypeConverter()->getIndexType(), args&: alignedPtr); |
| 2029 | return success(); |
| 2030 | } |
| 2031 | }; |
| 2032 | |
| 2033 | /// Materialize the MemRef descriptor represented by the results of |
| 2034 | /// ExtractStridedMetadataOp. |
| 2035 | class |
| 2036 | : public ConvertOpToLLVMPattern<memref::ExtractStridedMetadataOp> { |
| 2037 | public: |
| 2038 | using ConvertOpToLLVMPattern< |
| 2039 | memref::ExtractStridedMetadataOp>::ConvertOpToLLVMPattern; |
| 2040 | |
| 2041 | LogicalResult |
| 2042 | matchAndRewrite(memref::ExtractStridedMetadataOp , |
| 2043 | OpAdaptor adaptor, |
| 2044 | ConversionPatternRewriter &rewriter) const override { |
| 2045 | |
| 2046 | if (!LLVM::isCompatibleType(type: adaptor.getOperands().front().getType())) |
| 2047 | return failure(); |
| 2048 | |
| 2049 | // Create the descriptor. |
| 2050 | MemRefDescriptor sourceMemRef(adaptor.getSource()); |
| 2051 | Location loc = extractStridedMetadataOp.getLoc(); |
| 2052 | Value source = extractStridedMetadataOp.getSource(); |
| 2053 | |
| 2054 | auto sourceMemRefType = cast<MemRefType>(Val: source.getType()); |
| 2055 | int64_t rank = sourceMemRefType.getRank(); |
| 2056 | SmallVector<Value> results; |
| 2057 | results.reserve(N: 2 + rank * 2); |
| 2058 | |
| 2059 | // Base buffer. |
| 2060 | Value baseBuffer = sourceMemRef.allocatedPtr(builder&: rewriter, loc); |
| 2061 | Value alignedBuffer = sourceMemRef.alignedPtr(builder&: rewriter, loc); |
| 2062 | MemRefDescriptor dstMemRef = MemRefDescriptor::fromStaticShape( |
| 2063 | builder&: rewriter, loc, typeConverter: *getTypeConverter(), |
| 2064 | type: cast<MemRefType>(Val: extractStridedMetadataOp.getBaseBuffer().getType()), |
| 2065 | memory: baseBuffer, alignedMemory: alignedBuffer); |
| 2066 | results.push_back(Elt: (Value)dstMemRef); |
| 2067 | |
| 2068 | // Offset. |
| 2069 | results.push_back(Elt: sourceMemRef.offset(builder&: rewriter, loc)); |
| 2070 | |
| 2071 | // Sizes. |
| 2072 | for (unsigned i = 0; i < rank; ++i) |
| 2073 | results.push_back(Elt: sourceMemRef.size(builder&: rewriter, loc, pos: i)); |
| 2074 | // Strides. |
| 2075 | for (unsigned i = 0; i < rank; ++i) |
| 2076 | results.push_back(Elt: sourceMemRef.stride(builder&: rewriter, loc, pos: i)); |
| 2077 | |
| 2078 | rewriter.replaceOp(op: extractStridedMetadataOp, newValues: results); |
| 2079 | return success(); |
| 2080 | } |
| 2081 | }; |
| 2082 | |
| 2083 | } // namespace |
| 2084 | |
| 2085 | void mlir::populateFinalizeMemRefToLLVMConversionPatterns( |
| 2086 | const LLVMTypeConverter &converter, RewritePatternSet &patterns, |
| 2087 | SymbolTableCollection *symbolTables) { |
| 2088 | // clang-format off |
| 2089 | patterns.add< |
| 2090 | AllocaOpLowering, |
| 2091 | AllocaScopeOpLowering, |
| 2092 | AssumeAlignmentOpLowering, |
| 2093 | AtomicRMWOpLowering, |
| 2094 | ConvertExtractAlignedPointerAsIndex, |
| 2095 | DimOpLowering, |
| 2096 | DistinctObjectsOpLowering, |
| 2097 | ExtractStridedMetadataOpLowering, |
| 2098 | GenericAtomicRMWOpLowering, |
| 2099 | GetGlobalMemrefOpLowering, |
| 2100 | LoadOpLowering, |
| 2101 | MemRefCastOpLowering, |
| 2102 | MemRefReinterpretCastOpLowering, |
| 2103 | MemRefReshapeOpLowering, |
| 2104 | MemorySpaceCastOpLowering, |
| 2105 | PrefetchOpLowering, |
| 2106 | RankOpLowering, |
| 2107 | ReassociatingReshapeOpConversion<memref::CollapseShapeOp>, |
| 2108 | ReassociatingReshapeOpConversion<memref::ExpandShapeOp>, |
| 2109 | StoreOpLowering, |
| 2110 | SubViewOpLowering, |
| 2111 | TransposeOpLowering, |
| 2112 | ViewOpLowering>(arg: converter); |
| 2113 | // clang-format on |
| 2114 | patterns.add<GlobalMemrefOpLowering, MemRefCopyOpLowering>(arg: converter, |
| 2115 | args&: symbolTables); |
| 2116 | auto allocLowering = converter.getOptions().allocLowering; |
| 2117 | if (allocLowering == LowerToLLVMOptions::AllocLowering::AlignedAlloc) |
| 2118 | patterns.add<AlignedAllocOpLowering, DeallocOpLowering>(arg: converter, |
| 2119 | args&: symbolTables); |
| 2120 | else if (allocLowering == LowerToLLVMOptions::AllocLowering::Malloc) |
| 2121 | patterns.add<AllocOpLowering, DeallocOpLowering>(arg: converter, args&: symbolTables); |
| 2122 | } |
| 2123 | |
| 2124 | namespace { |
| 2125 | struct FinalizeMemRefToLLVMConversionPass |
| 2126 | : public impl::FinalizeMemRefToLLVMConversionPassBase< |
| 2127 | FinalizeMemRefToLLVMConversionPass> { |
| 2128 | using FinalizeMemRefToLLVMConversionPassBase:: |
| 2129 | FinalizeMemRefToLLVMConversionPassBase; |
| 2130 | |
| 2131 | void runOnOperation() override { |
| 2132 | Operation *op = getOperation(); |
| 2133 | const auto &dataLayoutAnalysis = getAnalysis<DataLayoutAnalysis>(); |
| 2134 | LowerToLLVMOptions options(&getContext(), |
| 2135 | dataLayoutAnalysis.getAtOrAbove(operation: op)); |
| 2136 | options.allocLowering = |
| 2137 | (useAlignedAlloc ? LowerToLLVMOptions::AllocLowering::AlignedAlloc |
| 2138 | : LowerToLLVMOptions::AllocLowering::Malloc); |
| 2139 | |
| 2140 | options.useGenericFunctions = useGenericFunctions; |
| 2141 | |
| 2142 | if (indexBitwidth != kDeriveIndexBitwidthFromDataLayout) |
| 2143 | options.overrideIndexBitwidth(bitwidth: indexBitwidth); |
| 2144 | |
| 2145 | LLVMTypeConverter typeConverter(&getContext(), options, |
| 2146 | &dataLayoutAnalysis); |
| 2147 | RewritePatternSet patterns(&getContext()); |
| 2148 | SymbolTableCollection symbolTables; |
| 2149 | populateFinalizeMemRefToLLVMConversionPatterns(converter: typeConverter, patterns, |
| 2150 | symbolTables: &symbolTables); |
| 2151 | LLVMConversionTarget target(getContext()); |
| 2152 | target.addLegalOp<func::FuncOp>(); |
| 2153 | if (failed(Result: applyPartialConversion(op, target, patterns: std::move(patterns)))) |
| 2154 | signalPassFailure(); |
| 2155 | } |
| 2156 | }; |
| 2157 | |
| 2158 | /// Implement the interface to convert MemRef to LLVM. |
| 2159 | struct MemRefToLLVMDialectInterface : public ConvertToLLVMPatternInterface { |
| 2160 | MemRefToLLVMDialectInterface(Dialect *dialect) |
| 2161 | : ConvertToLLVMPatternInterface(dialect) {} |
| 2162 | |
| 2163 | void loadDependentDialects(MLIRContext *context) const final { |
| 2164 | context->loadDialect<LLVM::LLVMDialect>(); |
| 2165 | } |
| 2166 | |
| 2167 | /// Hook for derived dialect interface to provide conversion patterns |
| 2168 | /// and mark dialect legal for the conversion target. |
| 2169 | void populateConvertToLLVMConversionPatterns( |
| 2170 | ConversionTarget &target, LLVMTypeConverter &typeConverter, |
| 2171 | RewritePatternSet &patterns) const final { |
| 2172 | populateFinalizeMemRefToLLVMConversionPatterns(converter: typeConverter, patterns); |
| 2173 | } |
| 2174 | }; |
| 2175 | |
| 2176 | } // namespace |
| 2177 | |
| 2178 | void mlir::registerConvertMemRefToLLVMInterface(DialectRegistry ®istry) { |
| 2179 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, memref::MemRefDialect *dialect) { |
| 2180 | dialect->addInterfaces<MemRefToLLVMDialectInterface>(); |
| 2181 | }); |
| 2182 | } |
| 2183 | |