| 1 | //===- ArithOps.cpp - MLIR Arith dialect ops implementation -----===// |
| 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 <cassert> |
| 10 | #include <cstdint> |
| 11 | #include <functional> |
| 12 | #include <type_traits> |
| 13 | #include <utility> |
| 14 | |
| 15 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 16 | #include "mlir/Dialect/CommonFolders.h" |
| 17 | #include "mlir/Dialect/UB/IR/UBMatchers.h" |
| 18 | #include "mlir/IR/Builders.h" |
| 19 | #include "mlir/IR/BuiltinAttributeInterfaces.h" |
| 20 | #include "mlir/IR/BuiltinAttributes.h" |
| 21 | #include "mlir/IR/Matchers.h" |
| 22 | #include "mlir/IR/OpImplementation.h" |
| 23 | #include "mlir/IR/PatternMatch.h" |
| 24 | #include "mlir/IR/TypeUtilities.h" |
| 25 | #include "mlir/Interfaces/SideEffectInterfaces.h" |
| 26 | #include "mlir/Support/LogicalResult.h" |
| 27 | |
| 28 | #include "llvm/ADT/APFloat.h" |
| 29 | #include "llvm/ADT/APInt.h" |
| 30 | #include "llvm/ADT/APSInt.h" |
| 31 | #include "llvm/ADT/DenseMap.h" |
| 32 | #include "llvm/ADT/FloatingPointMode.h" |
| 33 | #include "llvm/ADT/STLExtras.h" |
| 34 | #include "llvm/ADT/SmallVector.h" |
| 35 | #include "llvm/ADT/TypeSwitch.h" |
| 36 | |
| 37 | using namespace mlir; |
| 38 | using namespace mlir::arith; |
| 39 | |
| 40 | /// Default rounding mode according to default LLVM floating-point environment. |
| 41 | static constexpr llvm::RoundingMode kDefaultRoundingMode = |
| 42 | llvm::RoundingMode::NearestTiesToEven; |
| 43 | |
| 44 | //===----------------------------------------------------------------------===// |
| 45 | // Pattern helpers |
| 46 | //===----------------------------------------------------------------------===// |
| 47 | |
| 48 | static IntegerAttr |
| 49 | applyToIntegerAttrs(PatternRewriter &builder, Value res, Attribute lhs, |
| 50 | Attribute rhs, |
| 51 | function_ref<APInt(const APInt &, const APInt &)> binFn) { |
| 52 | const APInt &lhsVal = llvm::cast<IntegerAttr>(Val&: lhs).getValue(); |
| 53 | const APInt &rhsVal = llvm::cast<IntegerAttr>(Val&: rhs).getValue(); |
| 54 | APInt value = binFn(lhsVal, rhsVal); |
| 55 | return IntegerAttr::get(type: res.getType(), value); |
| 56 | } |
| 57 | |
| 58 | static IntegerAttr addIntegerAttrs(PatternRewriter &builder, Value res, |
| 59 | Attribute lhs, Attribute rhs) { |
| 60 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::plus<APInt>()); |
| 61 | } |
| 62 | |
| 63 | static IntegerAttr subIntegerAttrs(PatternRewriter &builder, Value res, |
| 64 | Attribute lhs, Attribute rhs) { |
| 65 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::minus<APInt>()); |
| 66 | } |
| 67 | |
| 68 | static IntegerAttr mulIntegerAttrs(PatternRewriter &builder, Value res, |
| 69 | Attribute lhs, Attribute rhs) { |
| 70 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::multiplies<APInt>()); |
| 71 | } |
| 72 | |
| 73 | static IntegerAttr andIntegerAttrs(PatternRewriter &builder, Value res, |
| 74 | Attribute lhs, Attribute rhs) { |
| 75 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::bit_and<APInt>()); |
| 76 | } |
| 77 | |
| 78 | static IntegerAttr orIntegerAttrs(PatternRewriter &builder, Value res, |
| 79 | Attribute lhs, Attribute rhs) { |
| 80 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::bit_or<APInt>()); |
| 81 | } |
| 82 | |
| 83 | static IntegerAttr xorIntegerAttrs(PatternRewriter &builder, Value res, |
| 84 | Attribute lhs, Attribute rhs) { |
| 85 | return applyToIntegerAttrs(builder, res, lhs, rhs, binFn: std::bit_xor<APInt>()); |
| 86 | } |
| 87 | |
| 88 | // Intersect overflow flags. Callers must ensure the common flags remain valid |
| 89 | // after rewriting. |
| 90 | static IntegerOverflowFlagsAttr |
| 91 | mergeOverflowFlags(IntegerOverflowFlagsAttr val1, |
| 92 | IntegerOverflowFlagsAttr val2) { |
| 93 | return IntegerOverflowFlagsAttr::get(context: val1.getContext(), |
| 94 | value: val1.getValue() & val2.getValue()); |
| 95 | } |
| 96 | |
| 97 | using OverflowCheckedBinaryOp = APInt (APInt::*)(const APInt &, bool &) const; |
| 98 | |
| 99 | // For fixed-width integers, if combining the constants does not overflow, the |
| 100 | // reassociated expression has the same mathematical result as the original |
| 101 | // expression. Common source flags then rule out overflow in the new operation. |
| 102 | static IntegerOverflowFlagsAttr mergeOverflowFlagsForReassociation( |
| 103 | IntegerOverflowFlagsAttr val1, IntegerOverflowFlagsAttr val2, Attribute lhs, |
| 104 | Attribute rhs, OverflowCheckedBinaryOp signedOp, |
| 105 | OverflowCheckedBinaryOp unsignedOp) { |
| 106 | IntegerOverflowFlags flags = val1.getValue() & val2.getValue(); |
| 107 | if (flags == IntegerOverflowFlags::none) |
| 108 | return IntegerOverflowFlagsAttr::get(context: val1.getContext(), value: flags); |
| 109 | |
| 110 | auto lhsAttr = llvm::cast<IntegerAttr>(Val&: lhs); |
| 111 | auto rhsAttr = llvm::cast<IntegerAttr>(Val&: rhs); |
| 112 | // TODO: Handle index separately; its storage width is not its target width. |
| 113 | if (isa<IndexType>(Val: lhsAttr.getType())) |
| 114 | return IntegerOverflowFlagsAttr::get(context: val1.getContext(), value: flags); |
| 115 | |
| 116 | const APInt &lhsValue = lhsAttr.getValue(); |
| 117 | const APInt &rhsValue = rhsAttr.getValue(); |
| 118 | bool overflow; |
| 119 | if (bitEnumContainsAny(bits: flags, bit: IntegerOverflowFlags::nsw)) { |
| 120 | (void)(lhsValue.*signedOp)(rhsValue, overflow); |
| 121 | if (overflow) |
| 122 | flags = bitEnumClear(bits: flags, bit: IntegerOverflowFlags::nsw); |
| 123 | } |
| 124 | if (bitEnumContainsAny(bits: flags, bit: IntegerOverflowFlags::nuw)) { |
| 125 | (void)(lhsValue.*unsignedOp)(rhsValue, overflow); |
| 126 | if (overflow) |
| 127 | flags = bitEnumClear(bits: flags, bit: IntegerOverflowFlags::nuw); |
| 128 | } |
| 129 | return IntegerOverflowFlagsAttr::get(context: val1.getContext(), value: flags); |
| 130 | } |
| 131 | |
| 132 | /// Invert an integer comparison predicate. |
| 133 | arith::CmpIPredicate arith::invertPredicate(arith::CmpIPredicate pred) { |
| 134 | switch (pred) { |
| 135 | case arith::CmpIPredicate::eq: |
| 136 | return arith::CmpIPredicate::ne; |
| 137 | case arith::CmpIPredicate::ne: |
| 138 | return arith::CmpIPredicate::eq; |
| 139 | case arith::CmpIPredicate::slt: |
| 140 | return arith::CmpIPredicate::sge; |
| 141 | case arith::CmpIPredicate::sle: |
| 142 | return arith::CmpIPredicate::sgt; |
| 143 | case arith::CmpIPredicate::sgt: |
| 144 | return arith::CmpIPredicate::sle; |
| 145 | case arith::CmpIPredicate::sge: |
| 146 | return arith::CmpIPredicate::slt; |
| 147 | case arith::CmpIPredicate::ult: |
| 148 | return arith::CmpIPredicate::uge; |
| 149 | case arith::CmpIPredicate::ule: |
| 150 | return arith::CmpIPredicate::ugt; |
| 151 | case arith::CmpIPredicate::ugt: |
| 152 | return arith::CmpIPredicate::ule; |
| 153 | case arith::CmpIPredicate::uge: |
| 154 | return arith::CmpIPredicate::ult; |
| 155 | } |
| 156 | llvm_unreachable("unknown cmpi predicate kind" ); |
| 157 | } |
| 158 | |
| 159 | /// Equivalent to |
| 160 | /// convertRoundingModeToLLVM(convertArithRoundingModeToLLVM(roundingMode)). |
| 161 | /// |
| 162 | /// Not possible to implement as chain of calls as this would introduce a |
| 163 | /// circular dependency with MLIRArithAttrToLLVMConversion and make arith depend |
| 164 | /// on the LLVM dialect and on translation to LLVM. |
| 165 | static llvm::RoundingMode |
| 166 | convertArithRoundingModeToLLVMIR(std::optional<RoundingMode> roundingMode) { |
| 167 | if (!roundingMode) |
| 168 | return kDefaultRoundingMode; |
| 169 | switch (*roundingMode) { |
| 170 | case RoundingMode::downward: |
| 171 | return llvm::RoundingMode::TowardNegative; |
| 172 | case RoundingMode::to_nearest_away: |
| 173 | return llvm::RoundingMode::NearestTiesToAway; |
| 174 | case RoundingMode::to_nearest_even: |
| 175 | return llvm::RoundingMode::NearestTiesToEven; |
| 176 | case RoundingMode::toward_zero: |
| 177 | return llvm::RoundingMode::TowardZero; |
| 178 | case RoundingMode::upward: |
| 179 | return llvm::RoundingMode::TowardPositive; |
| 180 | } |
| 181 | llvm_unreachable("Unhandled rounding mode" ); |
| 182 | } |
| 183 | |
| 184 | static arith::CmpIPredicateAttr invertPredicate(arith::CmpIPredicateAttr pred) { |
| 185 | return arith::CmpIPredicateAttr::get(context: pred.getContext(), |
| 186 | val: invertPredicate(pred: pred.getValue())); |
| 187 | } |
| 188 | |
| 189 | static int64_t getScalarOrElementWidth(Type type) { |
| 190 | Type elemTy = getElementTypeOrSelf(type); |
| 191 | if (elemTy.isIntOrFloat()) |
| 192 | return elemTy.getIntOrFloatBitWidth(); |
| 193 | |
| 194 | return -1; |
| 195 | } |
| 196 | |
| 197 | static int64_t getScalarOrElementWidth(Value value) { |
| 198 | return getScalarOrElementWidth(type: value.getType()); |
| 199 | } |
| 200 | |
| 201 | static FailureOr<APInt> getIntOrSplatIntValue(Attribute attr) { |
| 202 | APInt value; |
| 203 | if (matchPattern(attr, pattern: m_ConstantInt(bind_value: &value))) |
| 204 | return value; |
| 205 | |
| 206 | return failure(); |
| 207 | } |
| 208 | |
| 209 | static Attribute getBoolAttribute(Type type, bool value) { |
| 210 | auto boolAttr = BoolAttr::get(context: type.getContext(), value); |
| 211 | ShapedType shapedType = dyn_cast_or_null<ShapedType>(Val&: type); |
| 212 | if (!shapedType) |
| 213 | return boolAttr; |
| 214 | // DenseElementsAttr requires a static shape. |
| 215 | if (!shapedType.hasStaticShape()) |
| 216 | return {}; |
| 217 | return DenseElementsAttr::get(type: shapedType, values: boolAttr); |
| 218 | } |
| 219 | |
| 220 | /// Return a scalar or splat integer attribute of `type` (an integer/index type |
| 221 | /// or a shaped type thereof) holding `value`. Returns a null attribute for |
| 222 | /// shaped types with a dynamic shape, so callers can bail out of folding. |
| 223 | static Attribute getIntegerAttrOfType(Type type, int64_t value) { |
| 224 | auto scalarAttr = IntegerAttr::get(type: getElementTypeOrSelf(type), value); |
| 225 | ShapedType shapedType = dyn_cast<ShapedType>(Val&: type); |
| 226 | if (!shapedType) |
| 227 | return scalarAttr; |
| 228 | if (!shapedType.hasStaticShape()) |
| 229 | return {}; |
| 230 | return DenseElementsAttr::get(type: shapedType, values: scalarAttr); |
| 231 | } |
| 232 | |
| 233 | //===----------------------------------------------------------------------===// |
| 234 | // TableGen'd canonicalization patterns |
| 235 | //===----------------------------------------------------------------------===// |
| 236 | |
| 237 | namespace { |
| 238 | #include "ArithCanonicalization.inc" |
| 239 | } // namespace |
| 240 | |
| 241 | //===----------------------------------------------------------------------===// |
| 242 | // Common helpers |
| 243 | //===----------------------------------------------------------------------===// |
| 244 | |
| 245 | /// Return the type of the same shape (scalar, vector or tensor) containing i1. |
| 246 | static Type getI1SameShape(Type type) { |
| 247 | auto i1Type = IntegerType::get(context: type.getContext(), width: 1); |
| 248 | if (auto shapedType = dyn_cast<ShapedType>(Val&: type)) |
| 249 | return shapedType.cloneWith(shape: std::nullopt, elementType: i1Type); |
| 250 | if (llvm::isa<UnrankedTensorType>(Val: type)) |
| 251 | return UnrankedTensorType::get(elementType: i1Type); |
| 252 | return i1Type; |
| 253 | } |
| 254 | |
| 255 | //===----------------------------------------------------------------------===// |
| 256 | // ConstantOp |
| 257 | //===----------------------------------------------------------------------===// |
| 258 | |
| 259 | void arith::ConstantOp::getAsmResultNames( |
| 260 | function_ref<void(Value, StringRef)> setNameFn) { |
| 261 | auto type = getType(); |
| 262 | if (auto intCst = dyn_cast<IntegerAttr>(Val: getValue())) { |
| 263 | auto intType = dyn_cast<IntegerType>(Val&: type); |
| 264 | |
| 265 | // Sugar i1 constants with 'true' and 'false'. |
| 266 | if (intType && intType.getWidth() == 1) |
| 267 | return setNameFn(getResult(), (intCst.getInt() ? "true" : "false" )); |
| 268 | |
| 269 | // Otherwise, build a complex name with the value and type. |
| 270 | SmallString<32> specialNameBuffer; |
| 271 | llvm::raw_svector_ostream specialName(specialNameBuffer); |
| 272 | specialName << 'c' << intCst.getValue(); |
| 273 | if (intType) |
| 274 | specialName << '_' << type; |
| 275 | setNameFn(getResult(), specialName.str()); |
| 276 | } else { |
| 277 | setNameFn(getResult(), "cst" ); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | /// TODO: disallow arith.constant to return anything other than signless integer |
| 282 | /// or float like. |
| 283 | LogicalResult arith::ConstantOp::verify() { |
| 284 | auto type = getType(); |
| 285 | // Integer values must be signless. |
| 286 | if (auto intType = dyn_cast<IntegerType>(Val: getElementTypeOrSelf(type)); |
| 287 | intType && !intType.isSignless()) |
| 288 | return emitOpError(message: "integer return type must be signless" ); |
| 289 | // Any float or elements attribute are acceptable. |
| 290 | if (!llvm::isa<IntegerAttr, FloatAttr, ElementsAttr>(Val: getValue())) { |
| 291 | return emitOpError( |
| 292 | message: "value must be an integer, float, or elements attribute" ); |
| 293 | } |
| 294 | |
| 295 | // Note, we could relax this for vectors with 1 scalable dim, e.g.: |
| 296 | // * arith.constant dense<[[3, 3], [1, 1]]> : vector<2 x [2] x i32> |
| 297 | // However, this would most likely require updating the lowerings to LLVM. |
| 298 | if (isa<ScalableVectorType>(Val: type) && !isa<SplatElementsAttr>(Val: getValue())) |
| 299 | return emitOpError( |
| 300 | message: "initializing scalable vectors with elements attribute is not supported" |
| 301 | " unless it's a vector splat" ); |
| 302 | return success(); |
| 303 | } |
| 304 | |
| 305 | bool arith::ConstantOp::isBuildableWith(Attribute value, Type type) { |
| 306 | // The value's type must be the same as the provided type. |
| 307 | auto typedAttr = dyn_cast<TypedAttr>(Val&: value); |
| 308 | if (!typedAttr || typedAttr.getType() != type) |
| 309 | return false; |
| 310 | // Integer values must be signless. |
| 311 | if (auto intType = dyn_cast<IntegerType>(Val: getElementTypeOrSelf(type))) { |
| 312 | if (!intType.isSignless()) |
| 313 | return false; |
| 314 | } |
| 315 | // Integer, float, and element attributes are buildable. |
| 316 | return llvm::isa<IntegerAttr, FloatAttr, ElementsAttr>(Val: value); |
| 317 | } |
| 318 | |
| 319 | ConstantOp arith::ConstantOp::materialize(OpBuilder &builder, Attribute value, |
| 320 | Type type, Location loc) { |
| 321 | if (isBuildableWith(value, type)) |
| 322 | return arith::ConstantOp::create(builder, location: loc, value: cast<TypedAttr>(Val&: value)); |
| 323 | return nullptr; |
| 324 | } |
| 325 | |
| 326 | OpFoldResult arith::ConstantOp::fold(FoldAdaptor adaptor) { return getValue(); } |
| 327 | |
| 328 | void arith::ConstantIntOp::build(OpBuilder &builder, OperationState &result, |
| 329 | int64_t value, unsigned width) { |
| 330 | auto type = builder.getIntegerType(width); |
| 331 | arith::ConstantOp::build(odsBuilder&: builder, odsState&: result, result: type, |
| 332 | value: builder.getIntegerAttr(type, value)); |
| 333 | } |
| 334 | |
| 335 | arith::ConstantIntOp arith::ConstantIntOp::create(OpBuilder &builder, |
| 336 | Location location, |
| 337 | int64_t value, |
| 338 | unsigned width) { |
| 339 | mlir::OperationState state(location, getOperationName()); |
| 340 | build(builder, result&: state, value, width); |
| 341 | auto result = dyn_cast<ConstantIntOp>(Val: builder.create(state)); |
| 342 | assert(result && "builder didn't return the right type" ); |
| 343 | return result; |
| 344 | } |
| 345 | |
| 346 | arith::ConstantIntOp arith::ConstantIntOp::create(ImplicitLocOpBuilder &builder, |
| 347 | int64_t value, |
| 348 | unsigned width) { |
| 349 | return create(builder, location: builder.getLoc(), value, width); |
| 350 | } |
| 351 | |
| 352 | void arith::ConstantIntOp::build(OpBuilder &builder, OperationState &result, |
| 353 | Type type, int64_t value) { |
| 354 | arith::ConstantOp::build(odsBuilder&: builder, odsState&: result, result: type, |
| 355 | value: builder.getIntegerAttr(type, value)); |
| 356 | } |
| 357 | |
| 358 | arith::ConstantIntOp arith::ConstantIntOp::create(OpBuilder &builder, |
| 359 | Location location, Type type, |
| 360 | int64_t value) { |
| 361 | mlir::OperationState state(location, getOperationName()); |
| 362 | build(builder, result&: state, type, value); |
| 363 | auto result = dyn_cast<ConstantIntOp>(Val: builder.create(state)); |
| 364 | assert(result && "builder didn't return the right type" ); |
| 365 | return result; |
| 366 | } |
| 367 | |
| 368 | arith::ConstantIntOp arith::ConstantIntOp::create(ImplicitLocOpBuilder &builder, |
| 369 | Type type, int64_t value) { |
| 370 | return create(builder, location: builder.getLoc(), type, value); |
| 371 | } |
| 372 | |
| 373 | void arith::ConstantIntOp::build(OpBuilder &builder, OperationState &result, |
| 374 | Type type, const APInt &value) { |
| 375 | arith::ConstantOp::build(odsBuilder&: builder, odsState&: result, result: type, |
| 376 | value: builder.getIntegerAttr(type, value)); |
| 377 | } |
| 378 | |
| 379 | arith::ConstantIntOp arith::ConstantIntOp::create(OpBuilder &builder, |
| 380 | Location location, Type type, |
| 381 | const APInt &value) { |
| 382 | mlir::OperationState state(location, getOperationName()); |
| 383 | build(builder, result&: state, type, value); |
| 384 | auto result = dyn_cast<ConstantIntOp>(Val: builder.create(state)); |
| 385 | assert(result && "builder didn't return the right type" ); |
| 386 | return result; |
| 387 | } |
| 388 | |
| 389 | arith::ConstantIntOp arith::ConstantIntOp::create(ImplicitLocOpBuilder &builder, |
| 390 | Type type, |
| 391 | const APInt &value) { |
| 392 | return create(builder, location: builder.getLoc(), type, value); |
| 393 | } |
| 394 | |
| 395 | bool arith::ConstantIntOp::classof(Operation *op) { |
| 396 | if (auto constOp = dyn_cast_or_null<arith::ConstantOp>(Val: op)) |
| 397 | return constOp.getType().isSignlessInteger(); |
| 398 | return false; |
| 399 | } |
| 400 | |
| 401 | void arith::ConstantFloatOp::build(OpBuilder &builder, OperationState &result, |
| 402 | FloatType type, const APFloat &value) { |
| 403 | arith::ConstantOp::build(odsBuilder&: builder, odsState&: result, result: type, |
| 404 | value: builder.getFloatAttr(type, value)); |
| 405 | } |
| 406 | |
| 407 | arith::ConstantFloatOp arith::ConstantFloatOp::create(OpBuilder &builder, |
| 408 | Location location, |
| 409 | FloatType type, |
| 410 | const APFloat &value) { |
| 411 | mlir::OperationState state(location, getOperationName()); |
| 412 | build(builder, result&: state, type, value); |
| 413 | auto result = dyn_cast<ConstantFloatOp>(Val: builder.create(state)); |
| 414 | assert(result && "builder didn't return the right type" ); |
| 415 | return result; |
| 416 | } |
| 417 | |
| 418 | arith::ConstantFloatOp |
| 419 | arith::ConstantFloatOp::create(ImplicitLocOpBuilder &builder, FloatType type, |
| 420 | const APFloat &value) { |
| 421 | return create(builder, location: builder.getLoc(), type, value); |
| 422 | } |
| 423 | |
| 424 | bool arith::ConstantFloatOp::classof(Operation *op) { |
| 425 | if (auto constOp = dyn_cast_or_null<arith::ConstantOp>(Val: op)) |
| 426 | return llvm::isa<FloatType>(Val: constOp.getType()); |
| 427 | return false; |
| 428 | } |
| 429 | |
| 430 | void arith::ConstantIndexOp::build(OpBuilder &builder, OperationState &result, |
| 431 | int64_t value) { |
| 432 | arith::ConstantOp::build(odsBuilder&: builder, odsState&: result, result: builder.getIndexType(), |
| 433 | value: builder.getIndexAttr(value)); |
| 434 | } |
| 435 | |
| 436 | arith::ConstantIndexOp arith::ConstantIndexOp::create(OpBuilder &builder, |
| 437 | Location location, |
| 438 | int64_t value) { |
| 439 | mlir::OperationState state(location, getOperationName()); |
| 440 | build(builder, result&: state, value); |
| 441 | auto result = dyn_cast<ConstantIndexOp>(Val: builder.create(state)); |
| 442 | assert(result && "builder didn't return the right type" ); |
| 443 | return result; |
| 444 | } |
| 445 | |
| 446 | arith::ConstantIndexOp |
| 447 | arith::ConstantIndexOp::create(ImplicitLocOpBuilder &builder, int64_t value) { |
| 448 | return create(builder, location: builder.getLoc(), value); |
| 449 | } |
| 450 | |
| 451 | bool arith::ConstantIndexOp::classof(Operation *op) { |
| 452 | if (auto constOp = dyn_cast_or_null<arith::ConstantOp>(Val: op)) |
| 453 | return constOp.getType().isIndex(); |
| 454 | return false; |
| 455 | } |
| 456 | |
| 457 | Value mlir::arith::getZeroConstant(OpBuilder &builder, Location loc, |
| 458 | Type type) { |
| 459 | // TODO: Incorporate this check to `FloatAttr::get*`. |
| 460 | assert(!isa<Float8E8M0FNUType>(getElementTypeOrSelf(type)) && |
| 461 | "type doesn't have a zero representation" ); |
| 462 | TypedAttr zeroAttr = builder.getZeroAttr(type); |
| 463 | assert(zeroAttr && "unsupported type for zero attribute" ); |
| 464 | return arith::ConstantOp::create(builder, location: loc, value: zeroAttr); |
| 465 | } |
| 466 | |
| 467 | //===----------------------------------------------------------------------===// |
| 468 | // AddIOp |
| 469 | //===----------------------------------------------------------------------===// |
| 470 | |
| 471 | OpFoldResult arith::AddIOp::fold(FoldAdaptor adaptor) { |
| 472 | // addi(x, 0) -> x |
| 473 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 474 | return getLhs(); |
| 475 | |
| 476 | // addi(subi(a, b), b) -> a |
| 477 | if (auto sub = getLhs().getDefiningOp<SubIOp>()) |
| 478 | if (getRhs() == sub.getRhs()) |
| 479 | return sub.getLhs(); |
| 480 | |
| 481 | // addi(b, subi(a, b)) -> a |
| 482 | if (auto sub = getRhs().getDefiningOp<SubIOp>()) |
| 483 | if (getLhs() == sub.getRhs()) |
| 484 | return sub.getLhs(); |
| 485 | |
| 486 | return constFoldBinaryOp<IntegerAttr>( |
| 487 | operands: adaptor.getOperands(), |
| 488 | calculate: [](APInt a, const APInt &b) { return std::move(a) + b; }); |
| 489 | } |
| 490 | |
| 491 | void arith::AddIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 492 | MLIRContext *context) { |
| 493 | patterns.add<AddIAddConstant, AddISubConstantRHS, AddISubConstantLHS, |
| 494 | AddIMulNegativeOneRhs, AddIMulNegativeOneLhs>(arg&: context); |
| 495 | } |
| 496 | |
| 497 | //===----------------------------------------------------------------------===// |
| 498 | // AddUIExtendedOp |
| 499 | //===----------------------------------------------------------------------===// |
| 500 | |
| 501 | std::optional<SmallVector<int64_t, 4>> |
| 502 | arith::AddUIExtendedOp::getIterationShapeToUnroll() { |
| 503 | if (auto vt = dyn_cast<VectorType>(Val: getType(i: 0))) |
| 504 | return llvm::to_vector<4>(Range: vt.getShape()); |
| 505 | return std::nullopt; |
| 506 | } |
| 507 | |
| 508 | // Returns the overflow bit, assuming that `sum` is the result of unsigned |
| 509 | // addition of `operand` and another number. |
| 510 | static APInt calculateUnsignedOverflow(const APInt &sum, const APInt &operand) { |
| 511 | return sum.ult(RHS: operand) ? APInt::getAllOnes(numBits: 1) : APInt::getZero(numBits: 1); |
| 512 | } |
| 513 | |
| 514 | LogicalResult |
| 515 | arith::AddUIExtendedOp::fold(FoldAdaptor adaptor, |
| 516 | SmallVectorImpl<OpFoldResult> &results) { |
| 517 | Type overflowTy = getOverflow().getType(); |
| 518 | // addui_extended(x, 0) -> x, false |
| 519 | if (matchPattern(value: getRhs(), pattern: m_Zero())) { |
| 520 | Builder builder(getContext()); |
| 521 | auto falseValue = builder.getZeroAttr(type: overflowTy); |
| 522 | |
| 523 | results.push_back(Elt: getLhs()); |
| 524 | results.push_back(Elt: falseValue); |
| 525 | return success(); |
| 526 | } |
| 527 | |
| 528 | // addui_extended(constant_a, constant_b) -> constant_sum, constant_carry |
| 529 | // Let the `constFoldBinaryOp` utility attempt to fold the sum of both |
| 530 | // operands. If that succeeds, calculate the overflow bit based on the sum |
| 531 | // and the first (constant) operand, `lhs`. |
| 532 | if (Attribute sumAttr = constFoldBinaryOp<IntegerAttr>( |
| 533 | operands: adaptor.getOperands(), |
| 534 | calculate: [](APInt a, const APInt &b) { return std::move(a) + b; })) { |
| 535 | // If any operand is poison, propagate poison to both results. |
| 536 | if (matchPattern(attr: sumAttr, pattern: ub::m_Poison())) { |
| 537 | results.push_back(Elt: sumAttr); |
| 538 | results.push_back(Elt: sumAttr); |
| 539 | return success(); |
| 540 | } |
| 541 | Attribute overflowAttr = constFoldBinaryOp<IntegerAttr>( |
| 542 | operands: ArrayRef({sumAttr, adaptor.getLhs()}), |
| 543 | resultType: getI1SameShape(type: llvm::cast<TypedAttr>(Val&: sumAttr).getType()), |
| 544 | calculate&: calculateUnsignedOverflow); |
| 545 | if (!overflowAttr) |
| 546 | return failure(); |
| 547 | |
| 548 | results.push_back(Elt: sumAttr); |
| 549 | results.push_back(Elt: overflowAttr); |
| 550 | return success(); |
| 551 | } |
| 552 | |
| 553 | return failure(); |
| 554 | } |
| 555 | |
| 556 | void arith::AddUIExtendedOp::getCanonicalizationPatterns( |
| 557 | RewritePatternSet &patterns, MLIRContext *context) { |
| 558 | patterns.add<AddUIExtendedToAddI>(arg&: context); |
| 559 | } |
| 560 | |
| 561 | //===----------------------------------------------------------------------===// |
| 562 | // SubUIExtendedOp |
| 563 | //===----------------------------------------------------------------------===// |
| 564 | |
| 565 | std::optional<SmallVector<int64_t, 4>> |
| 566 | arith::SubUIExtendedOp::getIterationShapeToUnroll() { |
| 567 | if (auto vt = dyn_cast<VectorType>(Val: getType(i: 0))) |
| 568 | return llvm::to_vector<4>(Range: vt.getShape()); |
| 569 | return std::nullopt; |
| 570 | } |
| 571 | |
| 572 | // Returns the borrow bit, assuming `lhs` and `rhs` are operands of an unsigned |
| 573 | // subtraction whose mathematical result underflows iff `lhs < rhs`. |
| 574 | static APInt calculateUnsignedBorrow(const APInt &lhs, const APInt &rhs) { |
| 575 | return lhs.ult(RHS: rhs) ? APInt::getAllOnes(numBits: 1) : APInt::getZero(numBits: 1); |
| 576 | } |
| 577 | |
| 578 | LogicalResult |
| 579 | arith::SubUIExtendedOp::fold(FoldAdaptor adaptor, |
| 580 | SmallVectorImpl<OpFoldResult> &results) { |
| 581 | Type borrowTy = getBorrow().getType(); |
| 582 | // subui_extended(x, 0) -> x, false |
| 583 | if (matchPattern(value: getRhs(), pattern: m_Zero())) { |
| 584 | Builder builder(getContext()); |
| 585 | auto falseValue = builder.getZeroAttr(type: borrowTy); |
| 586 | |
| 587 | results.push_back(Elt: getLhs()); |
| 588 | results.push_back(Elt: falseValue); |
| 589 | return success(); |
| 590 | } |
| 591 | |
| 592 | // subui_extended(x, x) -> 0, false |
| 593 | if (getLhs() == getRhs()) { |
| 594 | // A dynamically-shaped result cannot be a constant; bail before |
| 595 | // getZeroAttr, which would assert on a non-static shape. |
| 596 | auto shapedType = dyn_cast<ShapedType>(Val: getDiff().getType()); |
| 597 | if (shapedType && !shapedType.hasStaticShape()) |
| 598 | return failure(); |
| 599 | Builder builder(getContext()); |
| 600 | auto zeroDiff = builder.getZeroAttr(type: getDiff().getType()); |
| 601 | auto falseValue = builder.getZeroAttr(type: borrowTy); |
| 602 | if (!zeroDiff) |
| 603 | return failure(); |
| 604 | |
| 605 | results.push_back(Elt: zeroDiff); |
| 606 | results.push_back(Elt: falseValue); |
| 607 | return success(); |
| 608 | } |
| 609 | |
| 610 | // subui_extended(constant_a, constant_b) -> constant_diff, constant_borrow |
| 611 | if (Attribute diffAttr = constFoldBinaryOp<IntegerAttr>( |
| 612 | operands: adaptor.getOperands(), |
| 613 | calculate: [](APInt a, const APInt &b) { return std::move(a) - b; })) { |
| 614 | // If any operand is poison, propagate poison to both results. |
| 615 | if (matchPattern(attr: diffAttr, pattern: ub::m_Poison())) { |
| 616 | results.push_back(Elt: diffAttr); |
| 617 | results.push_back(Elt: diffAttr); |
| 618 | return success(); |
| 619 | } |
| 620 | Attribute borrowAttr = constFoldBinaryOp<IntegerAttr>( |
| 621 | operands: adaptor.getOperands(), |
| 622 | resultType: getI1SameShape(type: llvm::cast<TypedAttr>(Val&: diffAttr).getType()), |
| 623 | calculate&: calculateUnsignedBorrow); |
| 624 | if (!borrowAttr) |
| 625 | return failure(); |
| 626 | |
| 627 | results.push_back(Elt: diffAttr); |
| 628 | results.push_back(Elt: borrowAttr); |
| 629 | return success(); |
| 630 | } |
| 631 | |
| 632 | return failure(); |
| 633 | } |
| 634 | |
| 635 | void arith::SubUIExtendedOp::getCanonicalizationPatterns( |
| 636 | RewritePatternSet &patterns, MLIRContext *context) { |
| 637 | patterns.add<SubUIExtendedToSubI>(arg&: context); |
| 638 | } |
| 639 | |
| 640 | //===----------------------------------------------------------------------===// |
| 641 | // SubIOp |
| 642 | //===----------------------------------------------------------------------===// |
| 643 | |
| 644 | OpFoldResult arith::SubIOp::fold(FoldAdaptor adaptor) { |
| 645 | // subi(x,x) -> 0 |
| 646 | if (getOperand(i: 0) == getOperand(i: 1)) { |
| 647 | auto shapedType = dyn_cast<ShapedType>(Val: getType()); |
| 648 | // We can't generate a constant with a dynamic shaped tensor. |
| 649 | if (!shapedType || shapedType.hasStaticShape()) |
| 650 | return Builder(getContext()).getZeroAttr(type: getType()); |
| 651 | } |
| 652 | // subi(x,0) -> x |
| 653 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 654 | return getLhs(); |
| 655 | |
| 656 | if (auto add = getLhs().getDefiningOp<AddIOp>()) { |
| 657 | // subi(addi(a, b), b) -> a |
| 658 | if (getRhs() == add.getRhs()) |
| 659 | return add.getLhs(); |
| 660 | // subi(addi(a, b), a) -> b |
| 661 | if (getRhs() == add.getLhs()) |
| 662 | return add.getRhs(); |
| 663 | } |
| 664 | |
| 665 | // subi(a, subi(a, b)) -> b |
| 666 | if (auto sub = getRhs().getDefiningOp<SubIOp>()) |
| 667 | if (getLhs() == sub.getLhs()) |
| 668 | return sub.getRhs(); |
| 669 | |
| 670 | return constFoldBinaryOp<IntegerAttr>( |
| 671 | operands: adaptor.getOperands(), |
| 672 | calculate: [](APInt a, const APInt &b) { return std::move(a) - b; }); |
| 673 | } |
| 674 | |
| 675 | void arith::SubIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 676 | MLIRContext *context) { |
| 677 | patterns.add<SubIRHSAddConstant, SubILHSAddConstant, SubIRHSSubConstantRHS, |
| 678 | SubIRHSSubConstantLHS, SubILHSSubConstantRHS, |
| 679 | SubILHSSubConstantLHS, SubISubILHSRHSLHS>(arg&: context); |
| 680 | } |
| 681 | |
| 682 | //===----------------------------------------------------------------------===// |
| 683 | // MulIOp |
| 684 | //===----------------------------------------------------------------------===// |
| 685 | |
| 686 | OpFoldResult arith::MulIOp::fold(FoldAdaptor adaptor) { |
| 687 | // muli(x, 0) -> 0 |
| 688 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 689 | return getRhs(); |
| 690 | // muli(x, 1) -> x |
| 691 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 692 | return getLhs(); |
| 693 | // TODO: Handle the overflow case. |
| 694 | |
| 695 | // default folder |
| 696 | return constFoldBinaryOp<IntegerAttr>( |
| 697 | operands: adaptor.getOperands(), |
| 698 | calculate: [](const APInt &a, const APInt &b) { return a * b; }); |
| 699 | } |
| 700 | |
| 701 | void arith::MulIOp::getAsmResultNames( |
| 702 | function_ref<void(Value, StringRef)> setNameFn) { |
| 703 | if (!isa<IndexType>(Val: getType())) |
| 704 | return; |
| 705 | |
| 706 | // Match vector.vscale by name to avoid depending on the vector dialect (which |
| 707 | // is a circular dependency). |
| 708 | auto isVscale = [](Operation *op) { |
| 709 | return op && op->getName().getStringRef() == "vector.vscale" ; |
| 710 | }; |
| 711 | |
| 712 | IntegerAttr baseValue; |
| 713 | auto isVscaleExpr = [&](Value a, Value b) { |
| 714 | return matchPattern(value: a, pattern: m_Constant(bind_value: &baseValue)) && |
| 715 | isVscale(b.getDefiningOp()); |
| 716 | }; |
| 717 | |
| 718 | if (!isVscaleExpr(getLhs(), getRhs()) && !isVscaleExpr(getRhs(), getLhs())) |
| 719 | return; |
| 720 | |
| 721 | // Name `base * vscale` or `vscale * base` as `c<base_value>_vscale`. |
| 722 | SmallString<32> specialNameBuffer; |
| 723 | llvm::raw_svector_ostream specialName(specialNameBuffer); |
| 724 | specialName << 'c' << baseValue.getInt() << "_vscale" ; |
| 725 | setNameFn(getResult(), specialName.str()); |
| 726 | } |
| 727 | |
| 728 | void arith::MulIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 729 | MLIRContext *context) { |
| 730 | patterns.add<MulIMulIConstant>(arg&: context); |
| 731 | } |
| 732 | |
| 733 | //===----------------------------------------------------------------------===// |
| 734 | // MulSIExtendedOp |
| 735 | //===----------------------------------------------------------------------===// |
| 736 | |
| 737 | std::optional<SmallVector<int64_t, 4>> |
| 738 | arith::MulSIExtendedOp::getIterationShapeToUnroll() { |
| 739 | if (auto vt = dyn_cast<VectorType>(Val: getType(i: 0))) |
| 740 | return llvm::to_vector<4>(Range: vt.getShape()); |
| 741 | return std::nullopt; |
| 742 | } |
| 743 | |
| 744 | LogicalResult |
| 745 | arith::MulSIExtendedOp::fold(FoldAdaptor adaptor, |
| 746 | SmallVectorImpl<OpFoldResult> &results) { |
| 747 | // mulsi_extended(x, 0) -> 0, 0 |
| 748 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) { |
| 749 | Attribute zero = adaptor.getRhs(); |
| 750 | results.push_back(Elt: zero); |
| 751 | results.push_back(Elt: zero); |
| 752 | return success(); |
| 753 | } |
| 754 | |
| 755 | // mulsi_extended(cst_a, cst_b) -> cst_low, cst_high |
| 756 | if (Attribute lowAttr = constFoldBinaryOp<IntegerAttr>( |
| 757 | operands: adaptor.getOperands(), |
| 758 | calculate: [](const APInt &a, const APInt &b) { return a * b; })) { |
| 759 | // Invoke the constant fold helper again to calculate the 'high' result. |
| 760 | Attribute highAttr = constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 761 | calculate&: llvm::APIntOps::mulhs); |
| 762 | assert(highAttr && "Unexpected constant-folding failure" ); |
| 763 | |
| 764 | results.push_back(Elt: lowAttr); |
| 765 | results.push_back(Elt: highAttr); |
| 766 | return success(); |
| 767 | } |
| 768 | |
| 769 | return failure(); |
| 770 | } |
| 771 | |
| 772 | void arith::MulSIExtendedOp::getCanonicalizationPatterns( |
| 773 | RewritePatternSet &patterns, MLIRContext *context) { |
| 774 | patterns.add<MulSIExtendedToMulI, MulSIExtendedRHSOne>(arg&: context); |
| 775 | } |
| 776 | |
| 777 | //===----------------------------------------------------------------------===// |
| 778 | // MulUIExtendedOp |
| 779 | //===----------------------------------------------------------------------===// |
| 780 | |
| 781 | std::optional<SmallVector<int64_t, 4>> |
| 782 | arith::MulUIExtendedOp::getIterationShapeToUnroll() { |
| 783 | if (auto vt = dyn_cast<VectorType>(Val: getType(i: 0))) |
| 784 | return llvm::to_vector<4>(Range: vt.getShape()); |
| 785 | return std::nullopt; |
| 786 | } |
| 787 | |
| 788 | LogicalResult |
| 789 | arith::MulUIExtendedOp::fold(FoldAdaptor adaptor, |
| 790 | SmallVectorImpl<OpFoldResult> &results) { |
| 791 | // mului_extended(x, 0) -> 0, 0 |
| 792 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) { |
| 793 | Attribute zero = adaptor.getRhs(); |
| 794 | results.push_back(Elt: zero); |
| 795 | results.push_back(Elt: zero); |
| 796 | return success(); |
| 797 | } |
| 798 | |
| 799 | // mului_extended(x, 1) -> x, 0 |
| 800 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) { |
| 801 | Builder builder(getContext()); |
| 802 | Attribute zero = builder.getZeroAttr(type: getLhs().getType()); |
| 803 | results.push_back(Elt: getLhs()); |
| 804 | results.push_back(Elt: zero); |
| 805 | return success(); |
| 806 | } |
| 807 | |
| 808 | // mului_extended(cst_a, cst_b) -> cst_low, cst_high |
| 809 | if (Attribute lowAttr = constFoldBinaryOp<IntegerAttr>( |
| 810 | operands: adaptor.getOperands(), |
| 811 | calculate: [](const APInt &a, const APInt &b) { return a * b; })) { |
| 812 | // Invoke the constant fold helper again to calculate the 'high' result. |
| 813 | Attribute highAttr = constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 814 | calculate&: llvm::APIntOps::mulhu); |
| 815 | assert(highAttr && "Unexpected constant-folding failure" ); |
| 816 | |
| 817 | results.push_back(Elt: lowAttr); |
| 818 | results.push_back(Elt: highAttr); |
| 819 | return success(); |
| 820 | } |
| 821 | |
| 822 | return failure(); |
| 823 | } |
| 824 | |
| 825 | void arith::MulUIExtendedOp::getCanonicalizationPatterns( |
| 826 | RewritePatternSet &patterns, MLIRContext *context) { |
| 827 | patterns.add<MulUIExtendedToMulI>(arg&: context); |
| 828 | } |
| 829 | |
| 830 | //===----------------------------------------------------------------------===// |
| 831 | // DivUIOp |
| 832 | //===----------------------------------------------------------------------===// |
| 833 | |
| 834 | /// Fold `(a * b) / b -> a` |
| 835 | static Value foldDivMul(Value lhs, Value rhs, |
| 836 | arith::IntegerOverflowFlags ovfFlags) { |
| 837 | auto mul = lhs.getDefiningOp<mlir::arith::MulIOp>(); |
| 838 | if (!mul || !bitEnumContainsAll(bits: mul.getOverflowFlags(), bit: ovfFlags)) |
| 839 | return {}; |
| 840 | |
| 841 | if (mul.getLhs() == rhs) |
| 842 | return mul.getRhs(); |
| 843 | |
| 844 | if (mul.getRhs() == rhs) |
| 845 | return mul.getLhs(); |
| 846 | |
| 847 | return {}; |
| 848 | } |
| 849 | |
| 850 | OpFoldResult arith::DivUIOp::fold(FoldAdaptor adaptor) { |
| 851 | // TODO: divui (x, 0) -> poison. Division by zero is undefined behaviour and |
| 852 | // could fold to poison, but that would make the arith dialect depend on the |
| 853 | // ub dialect to materialize ub.poison; left out for now. |
| 854 | |
| 855 | // divui (x, 1) -> x. |
| 856 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 857 | return getLhs(); |
| 858 | |
| 859 | // divui (0, x) -> 0. Division by zero is UB, so refining to 0 is valid. |
| 860 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 861 | return getLhs(); |
| 862 | |
| 863 | // divui (x, x) -> 1. |
| 864 | if (getLhs() == getRhs()) |
| 865 | return getIntegerAttrOfType(type: getType(), value: 1); |
| 866 | |
| 867 | // (a * b) / b -> a |
| 868 | if (Value val = foldDivMul(lhs: getLhs(), rhs: getRhs(), ovfFlags: IntegerOverflowFlags::nuw)) |
| 869 | return val; |
| 870 | |
| 871 | // Don't fold if it would require a division by zero. |
| 872 | bool div0 = false; |
| 873 | auto result = constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 874 | calculate: [&](APInt a, const APInt &b) { |
| 875 | if (div0 || !b) { |
| 876 | div0 = true; |
| 877 | return a; |
| 878 | } |
| 879 | return a.udiv(RHS: b); |
| 880 | }); |
| 881 | |
| 882 | return div0 ? Attribute() : result; |
| 883 | } |
| 884 | |
| 885 | /// Returns whether an unsigned division by `divisor` is speculatable. |
| 886 | static Speculation::Speculatability getDivUISpeculatability(Value divisor) { |
| 887 | // X / 0 => UB |
| 888 | if (matchPattern(value: divisor, pattern: m_IntRangeWithoutZeroU())) |
| 889 | return Speculation::Speculatable; |
| 890 | |
| 891 | return Speculation::NotSpeculatable; |
| 892 | } |
| 893 | |
| 894 | Speculation::Speculatability arith::DivUIOp::getSpeculatability() { |
| 895 | return getDivUISpeculatability(divisor: getRhs()); |
| 896 | } |
| 897 | |
| 898 | //===----------------------------------------------------------------------===// |
| 899 | // DivSIOp |
| 900 | //===----------------------------------------------------------------------===// |
| 901 | |
| 902 | OpFoldResult arith::DivSIOp::fold(FoldAdaptor adaptor) { |
| 903 | // TODO: divsi (x, 0) -> poison. Division by zero is undefined behaviour and |
| 904 | // could fold to poison, but that would make the arith dialect depend on the |
| 905 | // ub dialect to materialize ub.poison; left out for now. |
| 906 | |
| 907 | // divsi (x, 1) -> x. |
| 908 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 909 | return getLhs(); |
| 910 | |
| 911 | // divsi (0, x) -> 0. Division by zero is UB, so refining to 0 is valid. |
| 912 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 913 | return getLhs(); |
| 914 | |
| 915 | // divsi (x, x) -> 1. |
| 916 | if (getLhs() == getRhs()) |
| 917 | return getIntegerAttrOfType(type: getType(), value: 1); |
| 918 | |
| 919 | // (a * b) / b -> a |
| 920 | if (Value val = foldDivMul(lhs: getLhs(), rhs: getRhs(), ovfFlags: IntegerOverflowFlags::nsw)) |
| 921 | return val; |
| 922 | |
| 923 | // Don't fold if it would overflow or if it requires a division by zero. |
| 924 | bool overflowOrDiv0 = false; |
| 925 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 926 | operands: adaptor.getOperands(), calculate: [&](APInt a, const APInt &b) { |
| 927 | if (overflowOrDiv0 || !b) { |
| 928 | overflowOrDiv0 = true; |
| 929 | return a; |
| 930 | } |
| 931 | return a.sdiv_ov(RHS: b, Overflow&: overflowOrDiv0); |
| 932 | }); |
| 933 | |
| 934 | return overflowOrDiv0 ? Attribute() : result; |
| 935 | } |
| 936 | |
| 937 | /// Returns whether a signed division by `divisor` is speculatable. This |
| 938 | /// function conservatively assumes that all signed division by -1 are not |
| 939 | /// speculatable. |
| 940 | static Speculation::Speculatability getDivSISpeculatability(Value divisor) { |
| 941 | // X / 0 => UB |
| 942 | // INT_MIN / -1 => UB |
| 943 | if (matchPattern(value: divisor, pattern: m_IntRangeWithoutZeroS()) && |
| 944 | matchPattern(value: divisor, pattern: m_IntRangeWithoutNegOneS())) |
| 945 | return Speculation::Speculatable; |
| 946 | |
| 947 | return Speculation::NotSpeculatable; |
| 948 | } |
| 949 | |
| 950 | Speculation::Speculatability arith::DivSIOp::getSpeculatability() { |
| 951 | return getDivSISpeculatability(divisor: getRhs()); |
| 952 | } |
| 953 | |
| 954 | //===----------------------------------------------------------------------===// |
| 955 | // CeilDivUIOp |
| 956 | //===----------------------------------------------------------------------===// |
| 957 | |
| 958 | OpFoldResult arith::CeilDivUIOp::fold(FoldAdaptor adaptor) { |
| 959 | // TODO: ceildivui (x, 0) -> poison. Division by zero is undefined behaviour |
| 960 | // and could fold to poison, but that would make the arith dialect depend on |
| 961 | // the ub dialect to materialize ub.poison; left out for now. |
| 962 | |
| 963 | // ceildivui (x, 1) -> x. |
| 964 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 965 | return getLhs(); |
| 966 | |
| 967 | // ceildivui (0, x) -> 0. Division by zero is UB, so refining to 0 is valid. |
| 968 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 969 | return getLhs(); |
| 970 | |
| 971 | // ceildivui (x, x) -> 1. |
| 972 | if (getLhs() == getRhs()) |
| 973 | return getIntegerAttrOfType(type: getType(), value: 1); |
| 974 | |
| 975 | bool overflowOrDiv0 = false; |
| 976 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 977 | operands: adaptor.getOperands(), calculate: [&](APInt a, const APInt &b) { |
| 978 | if (overflowOrDiv0 || !b) { |
| 979 | overflowOrDiv0 = true; |
| 980 | return a; |
| 981 | } |
| 982 | APInt quotient = a.udiv(RHS: b); |
| 983 | if (!a.urem(RHS: b)) |
| 984 | return quotient; |
| 985 | APInt one(a.getBitWidth(), 1, true); |
| 986 | return quotient.uadd_ov(RHS: one, Overflow&: overflowOrDiv0); |
| 987 | }); |
| 988 | |
| 989 | return overflowOrDiv0 ? Attribute() : result; |
| 990 | } |
| 991 | |
| 992 | Speculation::Speculatability arith::CeilDivUIOp::getSpeculatability() { |
| 993 | return getDivUISpeculatability(divisor: getRhs()); |
| 994 | } |
| 995 | |
| 996 | //===----------------------------------------------------------------------===// |
| 997 | // CeilDivSIOp |
| 998 | //===----------------------------------------------------------------------===// |
| 999 | |
| 1000 | OpFoldResult arith::CeilDivSIOp::fold(FoldAdaptor adaptor) { |
| 1001 | // TODO: ceildivsi (x, 0) -> poison. Division by zero is undefined behaviour |
| 1002 | // and could fold to poison, but that would make the arith dialect depend on |
| 1003 | // the ub dialect to materialize ub.poison; left out for now. |
| 1004 | |
| 1005 | // ceildivsi (x, 1) -> x. |
| 1006 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 1007 | return getLhs(); |
| 1008 | |
| 1009 | // ceildivsi (0, x) -> 0. Division by zero is UB, so refining to 0 is valid. |
| 1010 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 1011 | return getLhs(); |
| 1012 | |
| 1013 | // ceildivsi (x, x) -> 1. |
| 1014 | if (getLhs() == getRhs()) |
| 1015 | return getIntegerAttrOfType(type: getType(), value: 1); |
| 1016 | |
| 1017 | // Don't fold if it would overflow or if it requires a division by zero. |
| 1018 | bool overflowOrDiv0 = false; |
| 1019 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 1020 | operands: adaptor.getOperands(), calculate: [&](const APInt &a, const APInt &b) { |
| 1021 | if (overflowOrDiv0 || !b) { |
| 1022 | overflowOrDiv0 = true; |
| 1023 | return a; |
| 1024 | } |
| 1025 | // Compute the ceiling without negating either operand, so that MININT |
| 1026 | // operands still fold whenever the result is representable. |
| 1027 | // |
| 1028 | // sdiv truncates towards zero, so it already rounds up whenever the |
| 1029 | // exact quotient is negative. When the exact quotient is positive, i.e. |
| 1030 | // when the operands have the same sign, an inexact division has to be |
| 1031 | // corrected by one. This mirrors the expansion in ExpandOps.cpp. |
| 1032 | bool overflowDiv = false; |
| 1033 | APInt quotient = a.sdiv_ov(RHS: b, Overflow&: overflowDiv); |
| 1034 | if (overflowDiv) { |
| 1035 | // MININT / -1. The exact result is -MININT, which is not |
| 1036 | // representable. |
| 1037 | overflowOrDiv0 = true; |
| 1038 | return a; |
| 1039 | } |
| 1040 | if (a.isNegative() != b.isNegative() || quotient * b == a) |
| 1041 | return quotient; |
| 1042 | |
| 1043 | // The correction cannot overflow: it only applies when the exact |
| 1044 | // quotient is positive and the division is inexact, which bounds the |
| 1045 | // quotient well below the maximum. Check anyway, at no cost. |
| 1046 | APInt one(a.getBitWidth(), 1, /*isSigned=*/true); |
| 1047 | return quotient.sadd_ov(RHS: one, Overflow&: overflowOrDiv0); |
| 1048 | }); |
| 1049 | |
| 1050 | return overflowOrDiv0 ? Attribute() : result; |
| 1051 | } |
| 1052 | |
| 1053 | Speculation::Speculatability arith::CeilDivSIOp::getSpeculatability() { |
| 1054 | return getDivSISpeculatability(divisor: getRhs()); |
| 1055 | } |
| 1056 | |
| 1057 | //===----------------------------------------------------------------------===// |
| 1058 | // FloorDivSIOp |
| 1059 | //===----------------------------------------------------------------------===// |
| 1060 | |
| 1061 | OpFoldResult arith::FloorDivSIOp::fold(FoldAdaptor adaptor) { |
| 1062 | // TODO: floordivsi (x, 0) -> poison. Division by zero is undefined behaviour |
| 1063 | // and could fold to poison, but that would make the arith dialect depend on |
| 1064 | // the ub dialect to materialize ub.poison; left out for now. |
| 1065 | |
| 1066 | // floordivsi (x, 1) -> x. |
| 1067 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 1068 | return getLhs(); |
| 1069 | |
| 1070 | // floordivsi (0, x) -> 0. Division by zero is UB, so refining to 0 is valid. |
| 1071 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 1072 | return getLhs(); |
| 1073 | |
| 1074 | // floordivsi (x, x) -> 1. |
| 1075 | if (getLhs() == getRhs()) |
| 1076 | return getIntegerAttrOfType(type: getType(), value: 1); |
| 1077 | |
| 1078 | // Don't fold if it would overflow or if it requires a division by zero. |
| 1079 | bool overflowOrDiv = false; |
| 1080 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 1081 | operands: adaptor.getOperands(), calculate: [&](APInt a, const APInt &b) { |
| 1082 | if (b.isZero()) { |
| 1083 | overflowOrDiv = true; |
| 1084 | return a; |
| 1085 | } |
| 1086 | return a.sfloordiv_ov(RHS: b, Overflow&: overflowOrDiv); |
| 1087 | }); |
| 1088 | |
| 1089 | return overflowOrDiv ? Attribute() : result; |
| 1090 | } |
| 1091 | |
| 1092 | Speculation::Speculatability FloorDivSIOp::getSpeculatability() { |
| 1093 | // Division by zero is UB. |
| 1094 | // Signed division overflow (INT_MIN / -1) is also UB. |
| 1095 | Value divisor = getRhs(); |
| 1096 | if (matchPattern(value: divisor, pattern: m_IntRangeWithoutZeroS()) && |
| 1097 | matchPattern(value: divisor, pattern: m_IntRangeWithoutNegOneS())) |
| 1098 | return Speculation::Speculatable; |
| 1099 | |
| 1100 | return Speculation::NotSpeculatable; |
| 1101 | } |
| 1102 | |
| 1103 | //===----------------------------------------------------------------------===// |
| 1104 | // RemUIOp |
| 1105 | //===----------------------------------------------------------------------===// |
| 1106 | |
| 1107 | OpFoldResult arith::RemUIOp::fold(FoldAdaptor adaptor) { |
| 1108 | // TODO: remui (x, 0) -> poison. Remainder by zero is undefined behaviour and |
| 1109 | // could fold to poison, but that would make the arith dialect depend on the |
| 1110 | // ub dialect to materialize ub.poison; left out for now. |
| 1111 | |
| 1112 | // remui (x, 1) -> 0. |
| 1113 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 1114 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 1115 | |
| 1116 | // remui (0, x) -> 0 and remui (x, x) -> 0. Division by zero is UB, so |
| 1117 | // refining to 0 is valid. |
| 1118 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero()) || getLhs() == getRhs()) |
| 1119 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 1120 | |
| 1121 | // Don't fold if it would require a division by zero. |
| 1122 | bool div0 = false; |
| 1123 | auto result = constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1124 | calculate: [&](APInt a, const APInt &b) { |
| 1125 | if (div0 || b.isZero()) { |
| 1126 | div0 = true; |
| 1127 | return a; |
| 1128 | } |
| 1129 | return a.urem(RHS: b); |
| 1130 | }); |
| 1131 | |
| 1132 | return div0 ? Attribute() : result; |
| 1133 | } |
| 1134 | |
| 1135 | Speculation::Speculatability arith::RemUIOp::getSpeculatability() { |
| 1136 | return getDivUISpeculatability(divisor: getRhs()); |
| 1137 | } |
| 1138 | |
| 1139 | //===----------------------------------------------------------------------===// |
| 1140 | // RemSIOp |
| 1141 | //===----------------------------------------------------------------------===// |
| 1142 | |
| 1143 | OpFoldResult arith::RemSIOp::fold(FoldAdaptor adaptor) { |
| 1144 | // TODO: remsi (x, 0) -> poison. Remainder by zero is undefined behaviour and |
| 1145 | // could fold to poison, but that would make the arith dialect depend on the |
| 1146 | // ub dialect to materialize ub.poison; left out for now. |
| 1147 | |
| 1148 | // remsi (x, 1) -> 0. |
| 1149 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) |
| 1150 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 1151 | |
| 1152 | // remsi (0, x) -> 0 and remsi (x, x) -> 0. Division by zero is UB, so |
| 1153 | // refining to 0 is valid. |
| 1154 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero()) || getLhs() == getRhs()) |
| 1155 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 1156 | |
| 1157 | // Don't fold if it would require a division by zero. |
| 1158 | bool div0 = false; |
| 1159 | auto result = constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1160 | calculate: [&](APInt a, const APInt &b) { |
| 1161 | if (div0 || b.isZero()) { |
| 1162 | div0 = true; |
| 1163 | return a; |
| 1164 | } |
| 1165 | return a.srem(RHS: b); |
| 1166 | }); |
| 1167 | |
| 1168 | return div0 ? Attribute() : result; |
| 1169 | } |
| 1170 | |
| 1171 | Speculation::Speculatability arith::RemSIOp::getSpeculatability() { |
| 1172 | // X % 0 => UB |
| 1173 | // X % -1 is well-defined (always 0), unlike X / -1 which can overflow. |
| 1174 | if (matchPattern(value: getRhs(), pattern: m_IntRangeWithoutZeroS())) |
| 1175 | return Speculation::Speculatable; |
| 1176 | |
| 1177 | return Speculation::NotSpeculatable; |
| 1178 | } |
| 1179 | |
| 1180 | //===----------------------------------------------------------------------===// |
| 1181 | // AndIOp |
| 1182 | //===----------------------------------------------------------------------===// |
| 1183 | |
| 1184 | /// Fold `op(a, op(a, b))` to `op(a, b)` for an associative, commutative and |
| 1185 | /// idempotent `op` (e.g. `and`, `or`). |
| 1186 | template <typename OpTy> |
| 1187 | static Value foldIdempotentOfSameOp(OpTy op) { |
| 1188 | for (bool reversePrev : {false, true}) { |
| 1189 | auto prev = (reversePrev ? op.getRhs() : op.getLhs()) |
| 1190 | .template getDefiningOp<OpTy>(); |
| 1191 | if (!prev) |
| 1192 | continue; |
| 1193 | |
| 1194 | Value other = (reversePrev ? op.getLhs() : op.getRhs()); |
| 1195 | if (other != prev.getLhs() && other != prev.getRhs()) |
| 1196 | continue; |
| 1197 | |
| 1198 | return prev.getResult(); |
| 1199 | } |
| 1200 | return {}; |
| 1201 | } |
| 1202 | |
| 1203 | OpFoldResult arith::AndIOp::fold(FoldAdaptor adaptor) { |
| 1204 | /// and(x, 0) -> 0 |
| 1205 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 1206 | return getRhs(); |
| 1207 | /// and(x, allOnes) -> x |
| 1208 | APInt intValue; |
| 1209 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &intValue)) && |
| 1210 | intValue.isAllOnes()) |
| 1211 | return getLhs(); |
| 1212 | /// and(x, not(x)) -> 0 |
| 1213 | if (matchPattern(value: getRhs(), pattern: m_Op<XOrIOp>(matchers: matchers::m_Val(v: getLhs()), |
| 1214 | matchers: m_ConstantInt(bind_value: &intValue))) && |
| 1215 | intValue.isAllOnes()) |
| 1216 | return Builder(getContext()).getZeroAttr(type: getType()); |
| 1217 | /// and(not(x), x) -> 0 |
| 1218 | if (matchPattern(value: getLhs(), pattern: m_Op<XOrIOp>(matchers: matchers::m_Val(v: getRhs()), |
| 1219 | matchers: m_ConstantInt(bind_value: &intValue))) && |
| 1220 | intValue.isAllOnes()) |
| 1221 | return Builder(getContext()).getZeroAttr(type: getType()); |
| 1222 | |
| 1223 | /// and(a, and(a, b)) -> and(a, b) |
| 1224 | if (Value result = foldIdempotentOfSameOp(op: *this)) |
| 1225 | return result; |
| 1226 | |
| 1227 | return constFoldBinaryOp<IntegerAttr>( |
| 1228 | operands: adaptor.getOperands(), |
| 1229 | calculate: [](APInt a, const APInt &b) { return std::move(a) & b; }); |
| 1230 | } |
| 1231 | |
| 1232 | //===----------------------------------------------------------------------===// |
| 1233 | // OrIOp |
| 1234 | //===----------------------------------------------------------------------===// |
| 1235 | |
| 1236 | OpFoldResult arith::OrIOp::fold(FoldAdaptor adaptor) { |
| 1237 | if (APInt rhsVal; matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &rhsVal))) { |
| 1238 | /// or(x, 0) -> x |
| 1239 | if (rhsVal.isZero()) |
| 1240 | return getLhs(); |
| 1241 | /// or(x, <all ones>) -> <all ones> |
| 1242 | if (rhsVal.isAllOnes()) |
| 1243 | return adaptor.getRhs(); |
| 1244 | } |
| 1245 | |
| 1246 | APInt intValue; |
| 1247 | /// or(x, xor(x, 1)) -> 1 |
| 1248 | if (matchPattern(value: getRhs(), pattern: m_Op<XOrIOp>(matchers: matchers::m_Val(v: getLhs()), |
| 1249 | matchers: m_ConstantInt(bind_value: &intValue))) && |
| 1250 | intValue.isAllOnes()) |
| 1251 | return getRhs().getDefiningOp<XOrIOp>().getRhs(); |
| 1252 | /// or(xor(x, 1), x) -> 1 |
| 1253 | if (matchPattern(value: getLhs(), pattern: m_Op<XOrIOp>(matchers: matchers::m_Val(v: getRhs()), |
| 1254 | matchers: m_ConstantInt(bind_value: &intValue))) && |
| 1255 | intValue.isAllOnes()) |
| 1256 | return getLhs().getDefiningOp<XOrIOp>().getRhs(); |
| 1257 | |
| 1258 | /// or(a, or(a, b)) -> or(a, b) |
| 1259 | if (Value result = foldIdempotentOfSameOp(op: *this)) |
| 1260 | return result; |
| 1261 | |
| 1262 | return constFoldBinaryOp<IntegerAttr>( |
| 1263 | operands: adaptor.getOperands(), |
| 1264 | calculate: [](APInt a, const APInt &b) { return std::move(a) | b; }); |
| 1265 | } |
| 1266 | |
| 1267 | //===----------------------------------------------------------------------===// |
| 1268 | // XOrIOp |
| 1269 | //===----------------------------------------------------------------------===// |
| 1270 | |
| 1271 | OpFoldResult arith::XOrIOp::fold(FoldAdaptor adaptor) { |
| 1272 | /// xor(x, 0) -> x |
| 1273 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 1274 | return getLhs(); |
| 1275 | /// xor(x, x) -> 0 |
| 1276 | if (getLhs() == getRhs()) { |
| 1277 | // A dynamically-shaped result cannot be a constant; bail before |
| 1278 | // getZeroAttr, which would assert on a non-static shape. |
| 1279 | auto shapedType = dyn_cast<ShapedType>(Val: getType()); |
| 1280 | if (!shapedType || shapedType.hasStaticShape()) |
| 1281 | return Builder(getContext()).getZeroAttr(type: getType()); |
| 1282 | } |
| 1283 | /// xor(xor(x, a), a) -> x |
| 1284 | /// xor(xor(a, x), a) -> x |
| 1285 | if (arith::XOrIOp prev = getLhs().getDefiningOp<arith::XOrIOp>()) { |
| 1286 | if (prev.getRhs() == getRhs()) |
| 1287 | return prev.getLhs(); |
| 1288 | if (prev.getLhs() == getRhs()) |
| 1289 | return prev.getRhs(); |
| 1290 | } |
| 1291 | /// xor(a, xor(x, a)) -> x |
| 1292 | /// xor(a, xor(a, x)) -> x |
| 1293 | if (arith::XOrIOp prev = getRhs().getDefiningOp<arith::XOrIOp>()) { |
| 1294 | if (prev.getRhs() == getLhs()) |
| 1295 | return prev.getLhs(); |
| 1296 | if (prev.getLhs() == getLhs()) |
| 1297 | return prev.getRhs(); |
| 1298 | } |
| 1299 | |
| 1300 | return constFoldBinaryOp<IntegerAttr>( |
| 1301 | operands: adaptor.getOperands(), |
| 1302 | calculate: [](APInt a, const APInt &b) { return std::move(a) ^ b; }); |
| 1303 | } |
| 1304 | |
| 1305 | void arith::XOrIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1306 | MLIRContext *context) { |
| 1307 | patterns.add<XOrIXOrIConstant, XOrINotCmpI, XOrIOfExtUI, XOrIOfExtSI>( |
| 1308 | arg&: context); |
| 1309 | } |
| 1310 | |
| 1311 | //===----------------------------------------------------------------------===// |
| 1312 | // NegFOp |
| 1313 | //===----------------------------------------------------------------------===// |
| 1314 | |
| 1315 | OpFoldResult arith::NegFOp::fold(FoldAdaptor adaptor) { |
| 1316 | /// negf(negf(x)) -> x |
| 1317 | if (auto op = this->getOperand().getDefiningOp<arith::NegFOp>()) |
| 1318 | return op.getOperand(); |
| 1319 | return constFoldUnaryOpConditional<FloatAttr>( |
| 1320 | operands: adaptor.getOperands(), calculate: [](const APFloat &a) -> std::optional<APFloat> { |
| 1321 | if (!APFloat::semanticsHasSignedRepr(a.getSemantics())) |
| 1322 | return std::nullopt; |
| 1323 | return -a; |
| 1324 | }); |
| 1325 | } |
| 1326 | |
| 1327 | //===----------------------------------------------------------------------===// |
| 1328 | // FlushDenormalsOp |
| 1329 | //===----------------------------------------------------------------------===// |
| 1330 | |
| 1331 | OpFoldResult arith::FlushDenormalsOp::fold(FoldAdaptor adaptor) { |
| 1332 | // TODO: Fold flush_denormals if the floating-point type does not support |
| 1333 | // denormals. There is currently no API to query this information from |
| 1334 | // APFloat. |
| 1335 | |
| 1336 | // flush_denormals(flush_denormals(x)) -> flush_denormals(x) |
| 1337 | if (auto op = this->getOperand().getDefiningOp<arith::FlushDenormalsOp>()) |
| 1338 | return op.getResult(); |
| 1339 | |
| 1340 | // Constant-fold flush_denormals if the operand is a constant. |
| 1341 | return constFoldUnaryOp<FloatAttr>( |
| 1342 | operands: adaptor.getOperands(), calculate: [](const APFloat &a) { |
| 1343 | if (a.isDenormal()) |
| 1344 | return APFloat::getZero(Sem: a.getSemantics(), Negative: a.isNegative()); |
| 1345 | return a; |
| 1346 | }); |
| 1347 | } |
| 1348 | |
| 1349 | //===----------------------------------------------------------------------===// |
| 1350 | // AddFOp |
| 1351 | //===----------------------------------------------------------------------===// |
| 1352 | |
| 1353 | OpFoldResult arith::AddFOp::fold(FoldAdaptor adaptor) { |
| 1354 | // addf(x, -0) -> x |
| 1355 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NegZeroFloat())) |
| 1356 | return getLhs(); |
| 1357 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_PosZeroFloat()) && |
| 1358 | bitEnumContainsAll(bits: adaptor.getFastmath(), bit: FastMathFlags::nsz)) |
| 1359 | return getLhs(); |
| 1360 | |
| 1361 | auto rm = getRoundingmode(); |
| 1362 | return constFoldBinaryOp<FloatAttr>( |
| 1363 | operands: adaptor.getOperands(), calculate: [rm](const APFloat &a, const APFloat &b) { |
| 1364 | APFloat result(a); |
| 1365 | result.add(RHS: b, RM: convertArithRoundingModeToLLVMIR(roundingMode: rm)); |
| 1366 | return result; |
| 1367 | }); |
| 1368 | } |
| 1369 | |
| 1370 | void arith::AddFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1371 | MLIRContext *context) { |
| 1372 | patterns.add<AddFOfNegFLhs, AddFOfNegFRhs>(arg&: context); |
| 1373 | } |
| 1374 | |
| 1375 | //===----------------------------------------------------------------------===// |
| 1376 | // SubFOp |
| 1377 | //===----------------------------------------------------------------------===// |
| 1378 | |
| 1379 | OpFoldResult arith::SubFOp::fold(FoldAdaptor adaptor) { |
| 1380 | // subf(x, +0) -> x |
| 1381 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_PosZeroFloat())) |
| 1382 | return getLhs(); |
| 1383 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NegZeroFloat()) && |
| 1384 | bitEnumContainsAll(bits: adaptor.getFastmath(), bit: FastMathFlags::nsz)) |
| 1385 | return getLhs(); |
| 1386 | |
| 1387 | auto rm = getRoundingmode(); |
| 1388 | return constFoldBinaryOpConditional<FloatAttr>( |
| 1389 | operands: adaptor.getOperands(), |
| 1390 | calculate: [rm](const APFloat &a, const APFloat &b) -> std::optional<APFloat> { |
| 1391 | // A negative difference is not representable in types such as |
| 1392 | // f8E8M0FNU, and APFloat::subtract does not support computing it. |
| 1393 | if (!APFloat::semanticsHasSignedRepr(a.getSemantics()) && |
| 1394 | a.compare(RHS: b) == APFloat::cmpLessThan) |
| 1395 | return std::nullopt; |
| 1396 | APFloat result(a); |
| 1397 | result.subtract(RHS: b, RM: convertArithRoundingModeToLLVMIR(roundingMode: rm)); |
| 1398 | return result; |
| 1399 | }); |
| 1400 | } |
| 1401 | |
| 1402 | void arith::SubFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1403 | MLIRContext *context) { |
| 1404 | patterns.add<SubFOfNegZero>(arg&: context); |
| 1405 | } |
| 1406 | |
| 1407 | namespace { |
| 1408 | |
| 1409 | /// Narrow an extremum whose operands were extended from the result type: |
| 1410 | /// |
| 1411 | /// trunc(extremum(ext(lhs), ext(rhs))) -> extremum(lhs, rhs) |
| 1412 | /// |
| 1413 | /// The concrete extension is part of the pattern so each extremum is only |
| 1414 | /// registered with extensions that preserve its ordering. |
| 1415 | /// For floating-point types, also require the extension to preserve every |
| 1416 | /// source value relevant under the extremum's fast-math flags. |
| 1417 | template <typename TruncOp, typename ExtOp, typename ExtremumOp> |
| 1418 | struct NarrowExtremum final : OpRewritePattern<TruncOp> { |
| 1419 | using OpRewritePattern<TruncOp>::OpRewritePattern; |
| 1420 | |
| 1421 | LogicalResult matchAndRewrite(TruncOp truncOp, |
| 1422 | PatternRewriter &rewriter) const override { |
| 1423 | auto extremumOp = truncOp.getIn().template getDefiningOp<ExtremumOp>(); |
| 1424 | if (!extremumOp || !extremumOp->hasOneUse()) |
| 1425 | return failure(); |
| 1426 | |
| 1427 | auto lhsExt = extremumOp.getLhs().template getDefiningOp<ExtOp>(); |
| 1428 | auto rhsExt = extremumOp.getRhs().template getDefiningOp<ExtOp>(); |
| 1429 | if (!lhsExt || !rhsExt) |
| 1430 | return failure(); |
| 1431 | |
| 1432 | Value lhs = lhsExt.getIn(); |
| 1433 | Value rhs = rhsExt.getIn(); |
| 1434 | Type narrowType = truncOp.getType(); |
| 1435 | if (lhs.getType() != narrowType || rhs.getType() != narrowType) |
| 1436 | return failure(); |
| 1437 | |
| 1438 | // A floating-point extension is not necessarily lossless between arbitrary |
| 1439 | // floating-point semantics, even when the destination has a larger bit |
| 1440 | // width. In particular, it may lose the sign of zero or quiet a signaling |
| 1441 | // NaN, either of which can change an extremum's result. `nnan` lets us |
| 1442 | // disregard NaN representation differences, but all other relevant source |
| 1443 | // values must be preserved. |
| 1444 | if (auto narrowFloatType = |
| 1445 | dyn_cast<FloatType>(Val: getElementTypeOrSelf(type: narrowType))) { |
| 1446 | auto wideFloatType = |
| 1447 | dyn_cast<FloatType>(getElementTypeOrSelf(extremumOp.getType())); |
| 1448 | if (!wideFloatType) |
| 1449 | return failure(); |
| 1450 | |
| 1451 | const llvm::fltSemantics &narrowSemantics = |
| 1452 | narrowFloatType.getFloatSemantics(); |
| 1453 | const llvm::fltSemantics &wideSemantics = |
| 1454 | wideFloatType.getFloatSemantics(); |
| 1455 | bool ignoreNaNs = false; |
| 1456 | if constexpr (std::is_same_v<TruncOp, TruncFOp>) |
| 1457 | ignoreNaNs = |
| 1458 | bitEnumContainsAll(extremumOp.getFastmath(), FastMathFlags::nnan); |
| 1459 | if (!llvm::APFloatBase::isLosslesslyConvertibleTo( |
| 1460 | From: narrowSemantics, To: wideSemantics, IgnoreNaNs: ignoreNaNs)) |
| 1461 | return failure(); |
| 1462 | } |
| 1463 | |
| 1464 | rewriter.replaceOpWithNewOp<ExtremumOp>( |
| 1465 | truncOp, TypeRange{narrowType}, ValueRange{lhs, rhs}, |
| 1466 | extremumOp.getProperties(), |
| 1467 | extremumOp->getDiscardableAttrDictionary().getValue()); |
| 1468 | return success(); |
| 1469 | } |
| 1470 | }; |
| 1471 | |
| 1472 | } // namespace |
| 1473 | |
| 1474 | //===----------------------------------------------------------------------===// |
| 1475 | // MaximumFOp |
| 1476 | //===----------------------------------------------------------------------===// |
| 1477 | |
| 1478 | OpFoldResult arith::MaximumFOp::fold(FoldAdaptor adaptor) { |
| 1479 | // maximumf(x,x) -> x |
| 1480 | if (getLhs() == getRhs()) |
| 1481 | return getRhs(); |
| 1482 | |
| 1483 | // maximumf(x, -inf) -> x |
| 1484 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NegInfFloat())) |
| 1485 | return getLhs(); |
| 1486 | |
| 1487 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::maximum); |
| 1488 | } |
| 1489 | |
| 1490 | //===----------------------------------------------------------------------===// |
| 1491 | // MaxNumFOp |
| 1492 | //===----------------------------------------------------------------------===// |
| 1493 | |
| 1494 | OpFoldResult arith::MaxNumFOp::fold(FoldAdaptor adaptor) { |
| 1495 | // maxnumf(x,x) -> x |
| 1496 | if (getLhs() == getRhs()) |
| 1497 | return getRhs(); |
| 1498 | |
| 1499 | // maxnumf(x, NaN) -> x |
| 1500 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NaNFloat())) |
| 1501 | return getLhs(); |
| 1502 | |
| 1503 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::maxnum); |
| 1504 | } |
| 1505 | |
| 1506 | //===----------------------------------------------------------------------===// |
| 1507 | // MaximumNumFOp |
| 1508 | //===----------------------------------------------------------------------===// |
| 1509 | |
| 1510 | OpFoldResult arith::MaximumNumFOp::fold(FoldAdaptor adaptor) { |
| 1511 | // maximumnumf(x,x) -> x |
| 1512 | if (getLhs() == getRhs()) |
| 1513 | return getRhs(); |
| 1514 | |
| 1515 | // maximumnumf(x, NaN) -> x |
| 1516 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NaNFloat())) |
| 1517 | return getLhs(); |
| 1518 | |
| 1519 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::maximumnum); |
| 1520 | } |
| 1521 | |
| 1522 | //===----------------------------------------------------------------------===// |
| 1523 | // MaxSIOp |
| 1524 | //===----------------------------------------------------------------------===// |
| 1525 | |
| 1526 | OpFoldResult MaxSIOp::fold(FoldAdaptor adaptor) { |
| 1527 | // maxsi(x,x) -> x |
| 1528 | if (getLhs() == getRhs()) |
| 1529 | return getRhs(); |
| 1530 | |
| 1531 | if (APInt intValue; |
| 1532 | matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &intValue))) { |
| 1533 | // maxsi(x,MAX_INT) -> MAX_INT |
| 1534 | if (intValue.isMaxSignedValue()) |
| 1535 | return getRhs(); |
| 1536 | // maxsi(x, MIN_INT) -> x |
| 1537 | if (intValue.isMinSignedValue()) |
| 1538 | return getLhs(); |
| 1539 | } |
| 1540 | |
| 1541 | return constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1542 | calculate&: llvm::APIntOps::smax); |
| 1543 | } |
| 1544 | |
| 1545 | //===----------------------------------------------------------------------===// |
| 1546 | // MaxUIOp |
| 1547 | //===----------------------------------------------------------------------===// |
| 1548 | |
| 1549 | OpFoldResult MaxUIOp::fold(FoldAdaptor adaptor) { |
| 1550 | // maxui(x,x) -> x |
| 1551 | if (getLhs() == getRhs()) |
| 1552 | return getRhs(); |
| 1553 | |
| 1554 | if (APInt intValue; |
| 1555 | matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &intValue))) { |
| 1556 | // maxui(x,MAX_INT) -> MAX_INT |
| 1557 | if (intValue.isMaxValue()) |
| 1558 | return getRhs(); |
| 1559 | // maxui(x, MIN_INT) -> x |
| 1560 | if (intValue.isMinValue()) |
| 1561 | return getLhs(); |
| 1562 | } |
| 1563 | |
| 1564 | return constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1565 | calculate&: llvm::APIntOps::umax); |
| 1566 | } |
| 1567 | |
| 1568 | //===----------------------------------------------------------------------===// |
| 1569 | // MinimumFOp |
| 1570 | //===----------------------------------------------------------------------===// |
| 1571 | |
| 1572 | OpFoldResult arith::MinimumFOp::fold(FoldAdaptor adaptor) { |
| 1573 | // minimumf(x,x) -> x |
| 1574 | if (getLhs() == getRhs()) |
| 1575 | return getRhs(); |
| 1576 | |
| 1577 | // minimumf(x, +inf) -> x |
| 1578 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_PosInfFloat())) |
| 1579 | return getLhs(); |
| 1580 | |
| 1581 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::minimum); |
| 1582 | } |
| 1583 | |
| 1584 | //===----------------------------------------------------------------------===// |
| 1585 | // MinNumFOp |
| 1586 | //===----------------------------------------------------------------------===// |
| 1587 | |
| 1588 | OpFoldResult arith::MinNumFOp::fold(FoldAdaptor adaptor) { |
| 1589 | // minnumf(x,x) -> x |
| 1590 | if (getLhs() == getRhs()) |
| 1591 | return getRhs(); |
| 1592 | |
| 1593 | // minnumf(x, NaN) -> x |
| 1594 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NaNFloat())) |
| 1595 | return getLhs(); |
| 1596 | |
| 1597 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::minnum); |
| 1598 | } |
| 1599 | |
| 1600 | //===----------------------------------------------------------------------===// |
| 1601 | // MinimumNumFOp |
| 1602 | //===----------------------------------------------------------------------===// |
| 1603 | |
| 1604 | OpFoldResult arith::MinimumNumFOp::fold(FoldAdaptor adaptor) { |
| 1605 | // minimumnumf(x,x) -> x |
| 1606 | if (getLhs() == getRhs()) |
| 1607 | return getRhs(); |
| 1608 | |
| 1609 | // minimumnumf(x, NaN) -> x |
| 1610 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_NaNFloat())) |
| 1611 | return getLhs(); |
| 1612 | |
| 1613 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), calculate&: llvm::minimumnum); |
| 1614 | } |
| 1615 | |
| 1616 | //===----------------------------------------------------------------------===// |
| 1617 | // MinSIOp |
| 1618 | //===----------------------------------------------------------------------===// |
| 1619 | |
| 1620 | OpFoldResult MinSIOp::fold(FoldAdaptor adaptor) { |
| 1621 | // minsi(x,x) -> x |
| 1622 | if (getLhs() == getRhs()) |
| 1623 | return getRhs(); |
| 1624 | |
| 1625 | if (APInt intValue; |
| 1626 | matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &intValue))) { |
| 1627 | // minsi(x,MIN_INT) -> MIN_INT |
| 1628 | if (intValue.isMinSignedValue()) |
| 1629 | return getRhs(); |
| 1630 | // minsi(x, MAX_INT) -> x |
| 1631 | if (intValue.isMaxSignedValue()) |
| 1632 | return getLhs(); |
| 1633 | } |
| 1634 | |
| 1635 | return constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1636 | calculate&: llvm::APIntOps::smin); |
| 1637 | } |
| 1638 | |
| 1639 | //===----------------------------------------------------------------------===// |
| 1640 | // MinUIOp |
| 1641 | //===----------------------------------------------------------------------===// |
| 1642 | |
| 1643 | OpFoldResult MinUIOp::fold(FoldAdaptor adaptor) { |
| 1644 | // minui(x,x) -> x |
| 1645 | if (getLhs() == getRhs()) |
| 1646 | return getRhs(); |
| 1647 | |
| 1648 | if (APInt intValue; |
| 1649 | matchPattern(attr: adaptor.getRhs(), pattern: m_ConstantInt(bind_value: &intValue))) { |
| 1650 | // minui(x,MIN_INT) -> MIN_INT |
| 1651 | if (intValue.isMinValue()) |
| 1652 | return getRhs(); |
| 1653 | // minui(x, MAX_INT) -> x |
| 1654 | if (intValue.isMaxValue()) |
| 1655 | return getLhs(); |
| 1656 | } |
| 1657 | |
| 1658 | return constFoldBinaryOp<IntegerAttr>(operands: adaptor.getOperands(), |
| 1659 | calculate&: llvm::APIntOps::umin); |
| 1660 | } |
| 1661 | |
| 1662 | //===----------------------------------------------------------------------===// |
| 1663 | // MulFOp |
| 1664 | //===----------------------------------------------------------------------===// |
| 1665 | |
| 1666 | OpFoldResult arith::MulFOp::fold(FoldAdaptor adaptor) { |
| 1667 | // mulf(x, 1) -> x |
| 1668 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_OneFloat())) |
| 1669 | return getLhs(); |
| 1670 | |
| 1671 | // Match LLVM InstSimplify: with nnan+nsz, X * 0 -> 0 for a non-constant X. |
| 1672 | // When both operands are constants, fall through to APFloat so IEEE signed |
| 1673 | // zeros are preserved (e.g. (-c) * +0.0 == -0.0). |
| 1674 | if (arith::bitEnumContainsAll(bits: getFastmath(), bit: arith::FastMathFlags::nnan | |
| 1675 | arith::FastMathFlags::nsz) && |
| 1676 | !(adaptor.getLhs() && adaptor.getRhs())) { |
| 1677 | // mulf(x, 0) -> 0 |
| 1678 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_AnyZeroFloat())) |
| 1679 | return getRhs(); |
| 1680 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_AnyZeroFloat())) |
| 1681 | return getLhs(); |
| 1682 | } |
| 1683 | |
| 1684 | auto rm = getRoundingmode(); |
| 1685 | return constFoldBinaryOp<FloatAttr>( |
| 1686 | operands: adaptor.getOperands(), calculate: [rm](const APFloat &a, const APFloat &b) { |
| 1687 | APFloat result(a); |
| 1688 | result.multiply(RHS: b, RM: convertArithRoundingModeToLLVMIR(roundingMode: rm)); |
| 1689 | return result; |
| 1690 | }); |
| 1691 | } |
| 1692 | |
| 1693 | void arith::MulFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1694 | MLIRContext *context) { |
| 1695 | patterns.add<MulFOfNegF>(arg&: context); |
| 1696 | } |
| 1697 | |
| 1698 | //===----------------------------------------------------------------------===// |
| 1699 | // DivFOp |
| 1700 | //===----------------------------------------------------------------------===// |
| 1701 | |
| 1702 | OpFoldResult arith::DivFOp::fold(FoldAdaptor adaptor) { |
| 1703 | // divf(x, 1) -> x |
| 1704 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_OneFloat())) |
| 1705 | return getLhs(); |
| 1706 | |
| 1707 | auto rm = getRoundingmode(); |
| 1708 | return constFoldBinaryOp<FloatAttr>( |
| 1709 | operands: adaptor.getOperands(), calculate: [rm](const APFloat &a, const APFloat &b) { |
| 1710 | APFloat result(a); |
| 1711 | result.divide(RHS: b, RM: convertArithRoundingModeToLLVMIR(roundingMode: rm)); |
| 1712 | return result; |
| 1713 | }); |
| 1714 | } |
| 1715 | |
| 1716 | void arith::DivFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1717 | MLIRContext *context) { |
| 1718 | patterns.add<DivFOfNegF>(arg&: context); |
| 1719 | } |
| 1720 | |
| 1721 | //===----------------------------------------------------------------------===// |
| 1722 | // RemFOp |
| 1723 | //===----------------------------------------------------------------------===// |
| 1724 | |
| 1725 | OpFoldResult arith::RemFOp::fold(FoldAdaptor adaptor) { |
| 1726 | return constFoldBinaryOp<FloatAttr>(operands: adaptor.getOperands(), |
| 1727 | calculate: [](const APFloat &a, const APFloat &b) { |
| 1728 | APFloat result(a); |
| 1729 | // APFloat::mod() offers the remainder |
| 1730 | // behavior we want, i.e. the result has |
| 1731 | // the sign of LHS operand. |
| 1732 | (void)result.mod(RHS: b); |
| 1733 | return result; |
| 1734 | }); |
| 1735 | } |
| 1736 | |
| 1737 | //===----------------------------------------------------------------------===// |
| 1738 | // Utility functions for verifying cast ops |
| 1739 | //===----------------------------------------------------------------------===// |
| 1740 | |
| 1741 | template <typename... Types> |
| 1742 | using type_list = std::tuple<Types...> *; |
| 1743 | |
| 1744 | /// Returns a non-null type only if the provided type is one of the allowed |
| 1745 | /// types or one of the allowed shaped types of the allowed types. Returns the |
| 1746 | /// element type if a valid shaped type is provided. |
| 1747 | template <typename... ShapedTypes, typename... ElementTypes> |
| 1748 | static Type getUnderlyingType(Type type, type_list<ShapedTypes...>, |
| 1749 | type_list<ElementTypes...>) { |
| 1750 | if (llvm::isa<ShapedType>(Val: type) && !llvm::isa<ShapedTypes...>(type)) |
| 1751 | return {}; |
| 1752 | |
| 1753 | auto underlyingType = getElementTypeOrSelf(type); |
| 1754 | if (!llvm::isa<ElementTypes...>(underlyingType)) |
| 1755 | return {}; |
| 1756 | |
| 1757 | return underlyingType; |
| 1758 | } |
| 1759 | |
| 1760 | /// Get allowed underlying types for vectors and tensors. |
| 1761 | template <typename... ElementTypes> |
| 1762 | static Type getTypeIfLike(Type type) { |
| 1763 | return getUnderlyingType(type, type_list<VectorType, TensorType>(), |
| 1764 | type_list<ElementTypes...>()); |
| 1765 | } |
| 1766 | |
| 1767 | /// Get allowed underlying types for vectors, tensors, and memrefs. |
| 1768 | template <typename... ElementTypes> |
| 1769 | static Type getTypeIfLikeOrMemRef(Type type) { |
| 1770 | return getUnderlyingType(type, |
| 1771 | type_list<VectorType, TensorType, MemRefType>(), |
| 1772 | type_list<ElementTypes...>()); |
| 1773 | } |
| 1774 | |
| 1775 | /// Return false if both types are ranked tensor with mismatching encoding. |
| 1776 | static bool hasSameEncoding(Type typeA, Type typeB) { |
| 1777 | auto rankedTensorA = dyn_cast<RankedTensorType>(Val&: typeA); |
| 1778 | auto rankedTensorB = dyn_cast<RankedTensorType>(Val&: typeB); |
| 1779 | if (!rankedTensorA || !rankedTensorB) |
| 1780 | return true; |
| 1781 | return rankedTensorA.getEncoding() == rankedTensorB.getEncoding(); |
| 1782 | } |
| 1783 | |
| 1784 | static bool areValidCastInputsAndOutputs(TypeRange inputs, TypeRange outputs) { |
| 1785 | if (inputs.size() != 1 || outputs.size() != 1) |
| 1786 | return false; |
| 1787 | if (!hasSameEncoding(typeA: inputs.front(), typeB: outputs.front())) |
| 1788 | return false; |
| 1789 | return succeeded(Result: verifyCompatibleShapes(types1: inputs.front(), types2: outputs.front())); |
| 1790 | } |
| 1791 | |
| 1792 | //===----------------------------------------------------------------------===// |
| 1793 | // Verifiers for integer and floating point extension/truncation ops |
| 1794 | //===----------------------------------------------------------------------===// |
| 1795 | |
| 1796 | // Extend ops can only extend to a wider type. |
| 1797 | template <typename ValType, typename Op> |
| 1798 | static LogicalResult verifyExtOp(Op op) { |
| 1799 | Type srcType = getElementTypeOrSelf(op.getIn().getType()); |
| 1800 | Type dstType = getElementTypeOrSelf(op.getType()); |
| 1801 | |
| 1802 | if (llvm::cast<ValType>(srcType).getWidth() >= |
| 1803 | llvm::cast<ValType>(dstType).getWidth()) |
| 1804 | return op.emitError("result type " ) |
| 1805 | << dstType << " must be wider than operand type " << srcType; |
| 1806 | |
| 1807 | return success(); |
| 1808 | } |
| 1809 | |
| 1810 | // Truncate ops can only truncate to a shorter type. |
| 1811 | template <typename ValType, typename Op> |
| 1812 | static LogicalResult verifyTruncateOp(Op op) { |
| 1813 | Type srcType = getElementTypeOrSelf(op.getIn().getType()); |
| 1814 | Type dstType = getElementTypeOrSelf(op.getType()); |
| 1815 | |
| 1816 | if (llvm::cast<ValType>(srcType).getWidth() <= |
| 1817 | llvm::cast<ValType>(dstType).getWidth()) |
| 1818 | return op.emitError("result type " ) |
| 1819 | << dstType << " must be shorter than operand type " << srcType; |
| 1820 | |
| 1821 | return success(); |
| 1822 | } |
| 1823 | |
| 1824 | /// Validate a cast that changes the width of a type. |
| 1825 | template <template <typename> class WidthComparator, typename... ElementTypes> |
| 1826 | static bool checkWidthChangeCast(TypeRange inputs, TypeRange outputs) { |
| 1827 | if (!areValidCastInputsAndOutputs(inputs, outputs)) |
| 1828 | return false; |
| 1829 | |
| 1830 | auto srcType = getTypeIfLike<ElementTypes...>(inputs.front()); |
| 1831 | auto dstType = getTypeIfLike<ElementTypes...>(outputs.front()); |
| 1832 | if (!srcType || !dstType) |
| 1833 | return false; |
| 1834 | |
| 1835 | return WidthComparator<unsigned>()(dstType.getIntOrFloatBitWidth(), |
| 1836 | srcType.getIntOrFloatBitWidth()); |
| 1837 | } |
| 1838 | |
| 1839 | /// Attempts to convert `sourceValue` to an APFloat value with |
| 1840 | /// `targetSemantics` and `roundingMode`, without any information loss. |
| 1841 | static FailureOr<APFloat> |
| 1842 | convertFloatValue(APFloat sourceValue, |
| 1843 | const llvm::fltSemantics &targetSemantics, |
| 1844 | llvm::RoundingMode roundingMode = kDefaultRoundingMode) { |
| 1845 | // Reject special values that are not representable in the target type before |
| 1846 | // calling APFloat::convert, which would llvm_unreachable on them. |
| 1847 | using fltNonfiniteBehavior = llvm::fltNonfiniteBehavior; |
| 1848 | if (sourceValue.isInfinity() && |
| 1849 | (targetSemantics.nonFiniteBehavior == fltNonfiniteBehavior::NanOnly || |
| 1850 | targetSemantics.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly)) |
| 1851 | return failure(); |
| 1852 | if (sourceValue.isNaN() && |
| 1853 | targetSemantics.nonFiniteBehavior == fltNonfiniteBehavior::FiniteOnly) |
| 1854 | return failure(); |
| 1855 | // A format without a sign (e.g. f8E8M0FNU) has no negative values, and |
| 1856 | // APFloat::convert reaches llvm_unreachable on a negative NaN or overflow. |
| 1857 | if (sourceValue.isNegative() && |
| 1858 | !APFloat::semanticsHasSignedRepr(targetSemantics)) |
| 1859 | return failure(); |
| 1860 | |
| 1861 | bool losesInfo = false; |
| 1862 | auto status = sourceValue.convert(ToSemantics: targetSemantics, RM: roundingMode, losesInfo: &losesInfo); |
| 1863 | if (losesInfo || status != APFloat::opOK) |
| 1864 | return failure(); |
| 1865 | |
| 1866 | return sourceValue; |
| 1867 | } |
| 1868 | |
| 1869 | //===----------------------------------------------------------------------===// |
| 1870 | // ExtUIOp |
| 1871 | //===----------------------------------------------------------------------===// |
| 1872 | |
| 1873 | OpFoldResult arith::ExtUIOp::fold(FoldAdaptor adaptor) { |
| 1874 | if (auto lhs = getIn().getDefiningOp<ExtUIOp>()) { |
| 1875 | // Only the inner extension's nneg speaks about the surviving source; the |
| 1876 | // outer flag described the already-extended value. |
| 1877 | setNonNeg(lhs.getNonNeg()); |
| 1878 | getInMutable().assign(value: lhs.getIn()); |
| 1879 | return getResult(); |
| 1880 | } |
| 1881 | |
| 1882 | Type resType = getElementTypeOrSelf(type: getType()); |
| 1883 | unsigned bitWidth = llvm::cast<IntegerType>(Val&: resType).getWidth(); |
| 1884 | return constFoldCastOp<IntegerAttr, IntegerAttr>( |
| 1885 | operands: adaptor.getOperands(), resType: getType(), |
| 1886 | calculate: [bitWidth](const APInt &a, bool &castStatus) { |
| 1887 | return a.zext(width: bitWidth); |
| 1888 | }); |
| 1889 | } |
| 1890 | |
| 1891 | bool arith::ExtUIOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 1892 | return checkWidthChangeCast<std::greater, IntegerType>(inputs, outputs); |
| 1893 | } |
| 1894 | |
| 1895 | LogicalResult arith::ExtUIOp::verify() { |
| 1896 | return verifyExtOp<IntegerType>(op: *this); |
| 1897 | } |
| 1898 | |
| 1899 | //===----------------------------------------------------------------------===// |
| 1900 | // ExtSIOp |
| 1901 | //===----------------------------------------------------------------------===// |
| 1902 | |
| 1903 | OpFoldResult arith::ExtSIOp::fold(FoldAdaptor adaptor) { |
| 1904 | if (auto lhs = getIn().getDefiningOp<ExtSIOp>()) { |
| 1905 | getInMutable().assign(value: lhs.getIn()); |
| 1906 | return getResult(); |
| 1907 | } |
| 1908 | |
| 1909 | Type resType = getElementTypeOrSelf(type: getType()); |
| 1910 | unsigned bitWidth = llvm::cast<IntegerType>(Val&: resType).getWidth(); |
| 1911 | return constFoldCastOp<IntegerAttr, IntegerAttr>( |
| 1912 | operands: adaptor.getOperands(), resType: getType(), |
| 1913 | calculate: [bitWidth](const APInt &a, bool &castStatus) { |
| 1914 | return a.sext(width: bitWidth); |
| 1915 | }); |
| 1916 | } |
| 1917 | |
| 1918 | bool arith::ExtSIOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 1919 | return checkWidthChangeCast<std::greater, IntegerType>(inputs, outputs); |
| 1920 | } |
| 1921 | |
| 1922 | void arith::ExtSIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 1923 | MLIRContext *context) { |
| 1924 | patterns.add<ExtSIOfExtUI>(arg&: context); |
| 1925 | } |
| 1926 | |
| 1927 | LogicalResult arith::ExtSIOp::verify() { |
| 1928 | return verifyExtOp<IntegerType>(op: *this); |
| 1929 | } |
| 1930 | |
| 1931 | //===----------------------------------------------------------------------===// |
| 1932 | // ExtFOp |
| 1933 | //===----------------------------------------------------------------------===// |
| 1934 | |
| 1935 | /// Fold extension of float constants when there is no information loss due the |
| 1936 | /// difference in fp semantics. |
| 1937 | OpFoldResult arith::ExtFOp::fold(FoldAdaptor adaptor) { |
| 1938 | if (auto truncFOp = getOperand().getDefiningOp<TruncFOp>()) { |
| 1939 | if (truncFOp.getOperand().getType() == getType()) { |
| 1940 | arith::FastMathFlags truncFMF = |
| 1941 | truncFOp.getFastmath().value_or(u: arith::FastMathFlags::none); |
| 1942 | bool isTruncContract = |
| 1943 | bitEnumContainsAll(bits: truncFMF, bit: arith::FastMathFlags::contract); |
| 1944 | arith::FastMathFlags extFMF = |
| 1945 | getFastmath().value_or(u: arith::FastMathFlags::none); |
| 1946 | bool isExtContract = |
| 1947 | bitEnumContainsAll(bits: extFMF, bit: arith::FastMathFlags::contract); |
| 1948 | if (isTruncContract && isExtContract) { |
| 1949 | return truncFOp.getOperand(); |
| 1950 | } |
| 1951 | } |
| 1952 | } |
| 1953 | |
| 1954 | auto resElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 1955 | const llvm::fltSemantics &targetSemantics = resElemType.getFloatSemantics(); |
| 1956 | return constFoldCastOp<FloatAttr, FloatAttr>( |
| 1957 | operands: adaptor.getOperands(), resType: getType(), |
| 1958 | calculate: [&targetSemantics](const APFloat &a, bool &castStatus) { |
| 1959 | FailureOr<APFloat> result = convertFloatValue(sourceValue: a, targetSemantics); |
| 1960 | if (failed(Result: result)) { |
| 1961 | castStatus = false; |
| 1962 | return a; |
| 1963 | } |
| 1964 | return *result; |
| 1965 | }); |
| 1966 | } |
| 1967 | |
| 1968 | bool arith::ExtFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 1969 | return checkWidthChangeCast<std::greater, FloatType>(inputs, outputs); |
| 1970 | } |
| 1971 | |
| 1972 | LogicalResult arith::ExtFOp::verify() { return verifyExtOp<FloatType>(op: *this); } |
| 1973 | |
| 1974 | //===----------------------------------------------------------------------===// |
| 1975 | // ScalingExtFOp |
| 1976 | //===----------------------------------------------------------------------===// |
| 1977 | |
| 1978 | /// Fold `calculate` element-wise over the operands of a scaling cast op. The |
| 1979 | /// `constFoldBinaryOp` helpers cannot be used: they bail out unless both |
| 1980 | /// operands have the same type, and `in` and `scale` never do. |
| 1981 | static Attribute foldScalingCastOp( |
| 1982 | Attribute inAttr, Attribute scaleAttr, Type resultType, |
| 1983 | function_ref<std::optional<APFloat>(const APFloat &, const APFloat &)> |
| 1984 | calculate) { |
| 1985 | // Poison propagates, as it does in the generic constant folders. |
| 1986 | if (isa_and_nonnull<ub::PoisonAttr>(Val: inAttr)) |
| 1987 | return inAttr; |
| 1988 | if (isa_and_nonnull<ub::PoisonAttr>(Val: scaleAttr)) |
| 1989 | return scaleAttr; |
| 1990 | |
| 1991 | if (!inAttr || !scaleAttr || !resultType) |
| 1992 | return {}; |
| 1993 | |
| 1994 | if (auto inFloat = dyn_cast<FloatAttr>(Val&: inAttr)) { |
| 1995 | auto scaleFloat = dyn_cast<FloatAttr>(Val&: scaleAttr); |
| 1996 | if (!scaleFloat) |
| 1997 | return {}; |
| 1998 | std::optional<APFloat> result = |
| 1999 | calculate(inFloat.getValue(), scaleFloat.getValue()); |
| 2000 | if (!result) |
| 2001 | return {}; |
| 2002 | return FloatAttr::get(type: resultType, value: *result); |
| 2003 | } |
| 2004 | |
| 2005 | auto inElements = dyn_cast<DenseFPElementsAttr>(Val&: inAttr); |
| 2006 | auto scaleElements = dyn_cast<DenseFPElementsAttr>(Val&: scaleAttr); |
| 2007 | auto shapedResultType = dyn_cast<ShapedType>(Val&: resultType); |
| 2008 | if (!inElements || !scaleElements || !shapedResultType || |
| 2009 | !shapedResultType.hasStaticShape() || |
| 2010 | inElements.getNumElements() != scaleElements.getNumElements()) |
| 2011 | return {}; |
| 2012 | |
| 2013 | // Both operands are splats, so avoid expanding the elements out. |
| 2014 | if (inElements.isSplat() && scaleElements.isSplat()) { |
| 2015 | std::optional<APFloat> result = |
| 2016 | calculate(inElements.getSplatValue<APFloat>(), |
| 2017 | scaleElements.getSplatValue<APFloat>()); |
| 2018 | if (!result) |
| 2019 | return {}; |
| 2020 | return DenseElementsAttr::get(type: shapedResultType, values: *result); |
| 2021 | } |
| 2022 | |
| 2023 | SmallVector<APFloat> results; |
| 2024 | results.reserve(N: inElements.getNumElements()); |
| 2025 | for (const auto &[in, scale] : llvm::zip_equal(t&: inElements, u&: scaleElements)) { |
| 2026 | std::optional<APFloat> result = calculate(in, scale); |
| 2027 | if (!result) |
| 2028 | return {}; |
| 2029 | results.push_back(Elt: *result); |
| 2030 | } |
| 2031 | return DenseElementsAttr::get(type: shapedResultType, values: results); |
| 2032 | } |
| 2033 | |
| 2034 | /// Only scales that already are f8E8M0FNU fold. What a wider scale means is |
| 2035 | /// unsettled -- the tree does not say whether truncating one to f8E8M0FNU |
| 2036 | /// rounds or takes its exponent -- so a folder should not settle it, see |
| 2037 | /// https://github.com/llvm/llvm-project/issues/215295. |
| 2038 | static bool isFoldableScalingScale(Value scale) { |
| 2039 | return isa<Float8E8M0FNUType>(Val: getElementTypeOrSelf(type: scale.getType())); |
| 2040 | } |
| 2041 | |
| 2042 | OpFoldResult arith::ScalingExtFOp::fold(FoldAdaptor adaptor) { |
| 2043 | // scaling_extf(in, scale) -> mulf(extf(in), extf(scale)), matching the |
| 2044 | // expansion in ExpandOps.cpp. As in arith.extf, the widening steps only fold |
| 2045 | // when they are lossless. |
| 2046 | if (!isFoldableScalingScale(scale: getScale())) |
| 2047 | return {}; |
| 2048 | |
| 2049 | auto resElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 2050 | const llvm::fltSemantics &resSemantics = resElemType.getFloatSemantics(); |
| 2051 | return foldScalingCastOp( |
| 2052 | inAttr: adaptor.getIn(), scaleAttr: adaptor.getScale(), resultType: getType(), |
| 2053 | calculate: [&resSemantics](const APFloat &in, |
| 2054 | const APFloat &scale) -> std::optional<APFloat> { |
| 2055 | FailureOr<APFloat> inExt = convertFloatValue(sourceValue: in, targetSemantics: resSemantics); |
| 2056 | FailureOr<APFloat> scaleExt = convertFloatValue(sourceValue: scale, targetSemantics: resSemantics); |
| 2057 | if (failed(Result: inExt) || failed(Result: scaleExt)) |
| 2058 | return std::nullopt; |
| 2059 | APFloat result(*inExt); |
| 2060 | result.multiply(RHS: *scaleExt, RM: kDefaultRoundingMode); |
| 2061 | return result; |
| 2062 | }); |
| 2063 | } |
| 2064 | |
| 2065 | bool arith::ScalingExtFOp::areCastCompatible(TypeRange inputs, |
| 2066 | TypeRange outputs) { |
| 2067 | return checkWidthChangeCast<std::greater, FloatType>(inputs: inputs.front(), outputs); |
| 2068 | } |
| 2069 | |
| 2070 | LogicalResult arith::ScalingExtFOp::verify() { |
| 2071 | return verifyExtOp<FloatType>(op: *this); |
| 2072 | } |
| 2073 | |
| 2074 | //===----------------------------------------------------------------------===// |
| 2075 | // TruncIOp |
| 2076 | //===----------------------------------------------------------------------===// |
| 2077 | |
| 2078 | OpFoldResult arith::TruncIOp::fold(FoldAdaptor adaptor) { |
| 2079 | if (matchPattern(value: getOperand(), pattern: m_Op<arith::ExtUIOp>()) || |
| 2080 | matchPattern(value: getOperand(), pattern: m_Op<arith::ExtSIOp>())) { |
| 2081 | Value src = getOperand().getDefiningOp()->getOperand(idx: 0); |
| 2082 | Type srcType = getElementTypeOrSelf(type: src.getType()); |
| 2083 | Type dstType = getElementTypeOrSelf(type: getType()); |
| 2084 | // trunci(zexti(a)) -> trunci(a) |
| 2085 | // trunci(sexti(a)) -> trunci(a) |
| 2086 | if (llvm::cast<IntegerType>(Val&: srcType).getWidth() > |
| 2087 | llvm::cast<IntegerType>(Val&: dstType).getWidth()) { |
| 2088 | setOperand(src); |
| 2089 | return getResult(); |
| 2090 | } |
| 2091 | |
| 2092 | // trunci(zexti(a)) -> a |
| 2093 | // trunci(sexti(a)) -> a |
| 2094 | if (srcType == dstType) |
| 2095 | return src; |
| 2096 | } |
| 2097 | |
| 2098 | // trunci(trunci(a)) -> trunci(a) |
| 2099 | if (auto truncOp = getOperand().getDefiningOp<arith::TruncIOp>()) { |
| 2100 | // Preserve a flag only if both truncations are lossless in that signedness. |
| 2101 | setOverflowFlags(getOverflowFlags() & truncOp.getOverflowFlags()); |
| 2102 | setOperand(truncOp.getIn()); |
| 2103 | return getResult(); |
| 2104 | } |
| 2105 | |
| 2106 | Type resType = getElementTypeOrSelf(type: getType()); |
| 2107 | unsigned bitWidth = llvm::cast<IntegerType>(Val&: resType).getWidth(); |
| 2108 | return constFoldCastOp<IntegerAttr, IntegerAttr>( |
| 2109 | operands: adaptor.getOperands(), resType: getType(), |
| 2110 | calculate: [bitWidth](const APInt &a, bool &castStatus) { |
| 2111 | return a.trunc(width: bitWidth); |
| 2112 | }); |
| 2113 | } |
| 2114 | |
| 2115 | bool arith::TruncIOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2116 | return checkWidthChangeCast<std::less, IntegerType>(inputs, outputs); |
| 2117 | } |
| 2118 | |
| 2119 | void arith::TruncIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2120 | MLIRContext *context) { |
| 2121 | patterns.add<NarrowExtremum<TruncIOp, ExtSIOp, MaxSIOp>, |
| 2122 | NarrowExtremum<TruncIOp, ExtSIOp, MinSIOp>, |
| 2123 | NarrowExtremum<TruncIOp, ExtUIOp, MaxUIOp>, |
| 2124 | NarrowExtremum<TruncIOp, ExtUIOp, MinUIOp>, TruncIExtSIToExtSI, |
| 2125 | TruncIExtUIToExtUI, TruncIShrSIToTrunciShrUI>(arg&: context); |
| 2126 | } |
| 2127 | |
| 2128 | LogicalResult arith::TruncIOp::verify() { |
| 2129 | return verifyTruncateOp<IntegerType>(op: *this); |
| 2130 | } |
| 2131 | |
| 2132 | //===----------------------------------------------------------------------===// |
| 2133 | // TruncFOp |
| 2134 | //===----------------------------------------------------------------------===// |
| 2135 | |
| 2136 | /// Perform safe const propagation for truncf, i.e., only propagate if FP value |
| 2137 | /// can be represented without precision loss. |
| 2138 | OpFoldResult arith::TruncFOp::fold(FoldAdaptor adaptor) { |
| 2139 | auto resElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 2140 | if (auto extOp = getOperand().getDefiningOp<arith::ExtFOp>()) { |
| 2141 | Value src = extOp.getIn(); |
| 2142 | auto srcType = cast<FloatType>(Val: getElementTypeOrSelf(type: src.getType())); |
| 2143 | auto intermediateType = |
| 2144 | cast<FloatType>(Val: getElementTypeOrSelf(type: extOp.getType())); |
| 2145 | // Check whether every source value round-trips through the intermediate |
| 2146 | // type, including signaling NaNs and signed zero. |
| 2147 | if (llvm::APFloatBase::isLosslesslyConvertibleTo( |
| 2148 | From: srcType.getFloatSemantics(), |
| 2149 | To: intermediateType.getFloatSemantics())) { |
| 2150 | // truncf(extf(a)) -> truncf(a) |
| 2151 | if (srcType.getWidth() > resElemType.getWidth()) { |
| 2152 | setOperand(src); |
| 2153 | return getResult(); |
| 2154 | } |
| 2155 | |
| 2156 | // truncf(extf(a)) -> a |
| 2157 | if (srcType == resElemType) |
| 2158 | return src; |
| 2159 | } |
| 2160 | } |
| 2161 | |
| 2162 | const llvm::fltSemantics &targetSemantics = resElemType.getFloatSemantics(); |
| 2163 | return constFoldCastOp<FloatAttr, FloatAttr>( |
| 2164 | operands: adaptor.getOperands(), resType: getType(), |
| 2165 | calculate: [this, &targetSemantics](const APFloat &a, bool &castStatus) { |
| 2166 | llvm::RoundingMode llvmRoundingMode = |
| 2167 | convertArithRoundingModeToLLVMIR(roundingMode: getRoundingmode()); |
| 2168 | FailureOr<APFloat> result = |
| 2169 | convertFloatValue(sourceValue: a, targetSemantics, roundingMode: llvmRoundingMode); |
| 2170 | if (failed(Result: result)) { |
| 2171 | castStatus = false; |
| 2172 | return a; |
| 2173 | } |
| 2174 | return *result; |
| 2175 | }); |
| 2176 | } |
| 2177 | |
| 2178 | void arith::TruncFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2179 | MLIRContext *context) { |
| 2180 | patterns.add<NarrowExtremum<TruncFOp, ExtFOp, MaximumFOp>, |
| 2181 | NarrowExtremum<TruncFOp, ExtFOp, MaxNumFOp>, |
| 2182 | NarrowExtremum<TruncFOp, ExtFOp, MaximumNumFOp>, |
| 2183 | NarrowExtremum<TruncFOp, ExtFOp, MinimumFOp>, |
| 2184 | NarrowExtremum<TruncFOp, ExtFOp, MinNumFOp>, |
| 2185 | NarrowExtremum<TruncFOp, ExtFOp, MinimumNumFOp>, |
| 2186 | TruncFSIToFPToSIToFP, TruncFUIToFPToUIToFP>(arg&: context); |
| 2187 | } |
| 2188 | |
| 2189 | bool arith::TruncFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2190 | return checkWidthChangeCast<std::less, FloatType>(inputs, outputs); |
| 2191 | } |
| 2192 | |
| 2193 | LogicalResult arith::TruncFOp::verify() { |
| 2194 | return verifyTruncateOp<FloatType>(op: *this); |
| 2195 | } |
| 2196 | |
| 2197 | //===----------------------------------------------------------------------===// |
| 2198 | // ConvertFOp |
| 2199 | //===----------------------------------------------------------------------===// |
| 2200 | |
| 2201 | OpFoldResult arith::ConvertFOp::fold(FoldAdaptor adaptor) { |
| 2202 | auto resElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 2203 | const llvm::fltSemantics &targetSemantics = resElemType.getFloatSemantics(); |
| 2204 | return constFoldCastOp<FloatAttr, FloatAttr>( |
| 2205 | operands: adaptor.getOperands(), resType: getType(), |
| 2206 | calculate: [this, &targetSemantics](const APFloat &a, bool &castStatus) { |
| 2207 | llvm::RoundingMode llvmRoundingMode = |
| 2208 | convertArithRoundingModeToLLVMIR(roundingMode: getRoundingmode()); |
| 2209 | FailureOr<APFloat> result = |
| 2210 | convertFloatValue(sourceValue: a, targetSemantics, roundingMode: llvmRoundingMode); |
| 2211 | if (failed(Result: result)) { |
| 2212 | castStatus = false; |
| 2213 | return a; |
| 2214 | } |
| 2215 | return *result; |
| 2216 | }); |
| 2217 | } |
| 2218 | |
| 2219 | bool arith::ConvertFOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2220 | if (!areValidCastInputsAndOutputs(inputs, outputs)) |
| 2221 | return false; |
| 2222 | auto srcType = getTypeIfLike<FloatType>(type: inputs.front()); |
| 2223 | auto dstType = getTypeIfLike<FloatType>(type: outputs.front()); |
| 2224 | if (!srcType || !dstType) |
| 2225 | return false; |
| 2226 | return srcType != dstType && |
| 2227 | srcType.getIntOrFloatBitWidth() == dstType.getIntOrFloatBitWidth(); |
| 2228 | } |
| 2229 | |
| 2230 | LogicalResult arith::ConvertFOp::verify() { |
| 2231 | auto srcType = cast<FloatType>(Val: getElementTypeOrSelf(type: getIn().getType())); |
| 2232 | auto dstType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 2233 | if (srcType == dstType) |
| 2234 | return emitError(message: "result element type " ) |
| 2235 | << dstType << " must be different from operand element type " |
| 2236 | << srcType; |
| 2237 | if (srcType.getWidth() != dstType.getWidth()) |
| 2238 | return emitError(message: "result element type " ) |
| 2239 | << dstType << " must have the same bitwidth as operand element type " |
| 2240 | << srcType; |
| 2241 | return success(); |
| 2242 | } |
| 2243 | |
| 2244 | //===----------------------------------------------------------------------===// |
| 2245 | // ScalingTruncFOp |
| 2246 | //===----------------------------------------------------------------------===// |
| 2247 | |
| 2248 | OpFoldResult arith::ScalingTruncFOp::fold(FoldAdaptor adaptor) { |
| 2249 | // scaling_truncf(in, scale) -> truncf(in / extf(scale)), matching the |
| 2250 | // expansion in ExpandOps.cpp. Unlike scaling_extf, the scale is widened to |
| 2251 | // the type of `in` rather than to the result type. |
| 2252 | if (!isFoldableScalingScale(scale: getScale())) |
| 2253 | return {}; |
| 2254 | |
| 2255 | auto inElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getIn().getType())); |
| 2256 | auto resElemType = cast<FloatType>(Val: getElementTypeOrSelf(type: getType())); |
| 2257 | const llvm::fltSemantics &inSemantics = inElemType.getFloatSemantics(); |
| 2258 | const llvm::fltSemantics &resSemantics = resElemType.getFloatSemantics(); |
| 2259 | llvm::RoundingMode roundingMode = |
| 2260 | convertArithRoundingModeToLLVMIR(roundingMode: getRoundingmode()); |
| 2261 | return foldScalingCastOp( |
| 2262 | inAttr: adaptor.getIn(), scaleAttr: adaptor.getScale(), resultType: getType(), |
| 2263 | calculate: [&](const APFloat &in, const APFloat &scale) -> std::optional<APFloat> { |
| 2264 | FailureOr<APFloat> scaleExt = convertFloatValue(sourceValue: scale, targetSemantics: inSemantics); |
| 2265 | if (failed(Result: scaleExt)) |
| 2266 | return std::nullopt; |
| 2267 | APFloat quotient(in); |
| 2268 | quotient.divide(RHS: *scaleExt, RM: kDefaultRoundingMode); |
| 2269 | FailureOr<APFloat> result = |
| 2270 | convertFloatValue(sourceValue: quotient, targetSemantics: resSemantics, roundingMode); |
| 2271 | if (failed(Result: result)) |
| 2272 | return std::nullopt; |
| 2273 | return *result; |
| 2274 | }); |
| 2275 | } |
| 2276 | |
| 2277 | bool arith::ScalingTruncFOp::areCastCompatible(TypeRange inputs, |
| 2278 | TypeRange outputs) { |
| 2279 | return checkWidthChangeCast<std::less, FloatType>(inputs: inputs.front(), outputs); |
| 2280 | } |
| 2281 | |
| 2282 | LogicalResult arith::ScalingTruncFOp::verify() { |
| 2283 | return verifyTruncateOp<FloatType>(op: *this); |
| 2284 | } |
| 2285 | |
| 2286 | //===----------------------------------------------------------------------===// |
| 2287 | // AndIOp |
| 2288 | //===----------------------------------------------------------------------===// |
| 2289 | |
| 2290 | void arith::AndIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2291 | MLIRContext *context) { |
| 2292 | patterns.add<AndIAndIConstant, AndOfExtUI, AndOfExtSI>(arg&: context); |
| 2293 | } |
| 2294 | |
| 2295 | //===----------------------------------------------------------------------===// |
| 2296 | // OrIOp |
| 2297 | //===----------------------------------------------------------------------===// |
| 2298 | |
| 2299 | void arith::OrIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2300 | MLIRContext *context) { |
| 2301 | patterns.add<OrIOrIConstant, OrOfExtUI, OrOfExtSI>(arg&: context); |
| 2302 | } |
| 2303 | |
| 2304 | //===----------------------------------------------------------------------===// |
| 2305 | // Verifiers for casts between integers and floats. |
| 2306 | //===----------------------------------------------------------------------===// |
| 2307 | |
| 2308 | template <typename From, typename To> |
| 2309 | static bool checkIntFloatCast(TypeRange inputs, TypeRange outputs) { |
| 2310 | if (!areValidCastInputsAndOutputs(inputs, outputs)) |
| 2311 | return false; |
| 2312 | |
| 2313 | auto srcType = getTypeIfLike<From>(inputs.front()); |
| 2314 | auto dstType = getTypeIfLike<To>(outputs.back()); |
| 2315 | |
| 2316 | return srcType && dstType; |
| 2317 | } |
| 2318 | |
| 2319 | //===----------------------------------------------------------------------===// |
| 2320 | // UIToFPOp |
| 2321 | //===----------------------------------------------------------------------===// |
| 2322 | |
| 2323 | bool arith::UIToFPOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2324 | return checkIntFloatCast<IntegerType, FloatType>(inputs, outputs); |
| 2325 | } |
| 2326 | |
| 2327 | OpFoldResult arith::UIToFPOp::fold(FoldAdaptor adaptor) { |
| 2328 | Type resEleType = getElementTypeOrSelf(type: getType()); |
| 2329 | return constFoldCastOp<IntegerAttr, FloatAttr>( |
| 2330 | operands: adaptor.getOperands(), resType: getType(), |
| 2331 | calculate: [&resEleType](const APInt &a, bool &castStatus) { |
| 2332 | FloatType floatTy = llvm::cast<FloatType>(Val&: resEleType); |
| 2333 | APFloat apf(floatTy.getFloatSemantics(), |
| 2334 | APInt::getZero(numBits: floatTy.getWidth())); |
| 2335 | apf.convertFromAPInt(Input: a, /*IsSigned=*/false, |
| 2336 | RM: APFloat::rmNearestTiesToEven); |
| 2337 | return apf; |
| 2338 | }); |
| 2339 | } |
| 2340 | |
| 2341 | void arith::UIToFPOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2342 | MLIRContext *context) { |
| 2343 | patterns.add<UIToFPOfExtUI>(arg&: context); |
| 2344 | } |
| 2345 | |
| 2346 | //===----------------------------------------------------------------------===// |
| 2347 | // SIToFPOp |
| 2348 | //===----------------------------------------------------------------------===// |
| 2349 | |
| 2350 | bool arith::SIToFPOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2351 | return checkIntFloatCast<IntegerType, FloatType>(inputs, outputs); |
| 2352 | } |
| 2353 | |
| 2354 | OpFoldResult arith::SIToFPOp::fold(FoldAdaptor adaptor) { |
| 2355 | Type resEleType = getElementTypeOrSelf(type: getType()); |
| 2356 | return constFoldCastOp<IntegerAttr, FloatAttr>( |
| 2357 | operands: adaptor.getOperands(), resType: getType(), |
| 2358 | calculate: [&resEleType](const APInt &a, bool &castStatus) { |
| 2359 | FloatType floatTy = llvm::cast<FloatType>(Val&: resEleType); |
| 2360 | APFloat apf(floatTy.getFloatSemantics(), |
| 2361 | APInt::getZero(numBits: floatTy.getWidth())); |
| 2362 | if (a.isNegative() && |
| 2363 | !APFloat::semanticsHasSignedRepr(floatTy.getFloatSemantics())) { |
| 2364 | castStatus = false; |
| 2365 | return apf; |
| 2366 | } |
| 2367 | apf.convertFromAPInt(Input: a, /*IsSigned=*/true, |
| 2368 | RM: APFloat::rmNearestTiesToEven); |
| 2369 | return apf; |
| 2370 | }); |
| 2371 | } |
| 2372 | |
| 2373 | void arith::SIToFPOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2374 | MLIRContext *context) { |
| 2375 | patterns.add<SIToFPOfExtSI, SIToFPOfExtUI>(arg&: context); |
| 2376 | } |
| 2377 | |
| 2378 | //===----------------------------------------------------------------------===// |
| 2379 | // FPToUIOp |
| 2380 | //===----------------------------------------------------------------------===// |
| 2381 | |
| 2382 | bool arith::FPToUIOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2383 | return checkIntFloatCast<FloatType, IntegerType>(inputs, outputs); |
| 2384 | } |
| 2385 | |
| 2386 | OpFoldResult arith::FPToUIOp::fold(FoldAdaptor adaptor) { |
| 2387 | Type resType = getElementTypeOrSelf(type: getType()); |
| 2388 | unsigned bitWidth = llvm::cast<IntegerType>(Val&: resType).getWidth(); |
| 2389 | return constFoldCastOp<FloatAttr, IntegerAttr>( |
| 2390 | operands: adaptor.getOperands(), resType: getType(), |
| 2391 | calculate: [&bitWidth](const APFloat &a, bool &castStatus) { |
| 2392 | bool ignored; |
| 2393 | APSInt api(bitWidth, /*isUnsigned=*/true); |
| 2394 | castStatus = APFloat::opInvalidOp != |
| 2395 | a.convertToInteger(Result&: api, RM: APFloat::rmTowardZero, IsExact: &ignored); |
| 2396 | return api; |
| 2397 | }); |
| 2398 | } |
| 2399 | |
| 2400 | //===----------------------------------------------------------------------===// |
| 2401 | // FPToSIOp |
| 2402 | //===----------------------------------------------------------------------===// |
| 2403 | |
| 2404 | bool arith::FPToSIOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2405 | return checkIntFloatCast<FloatType, IntegerType>(inputs, outputs); |
| 2406 | } |
| 2407 | |
| 2408 | OpFoldResult arith::FPToSIOp::fold(FoldAdaptor adaptor) { |
| 2409 | Type resType = getElementTypeOrSelf(type: getType()); |
| 2410 | unsigned bitWidth = llvm::cast<IntegerType>(Val&: resType).getWidth(); |
| 2411 | return constFoldCastOp<FloatAttr, IntegerAttr>( |
| 2412 | operands: adaptor.getOperands(), resType: getType(), |
| 2413 | calculate: [&bitWidth](const APFloat &a, bool &castStatus) { |
| 2414 | bool ignored; |
| 2415 | APSInt api(bitWidth, /*isUnsigned=*/false); |
| 2416 | castStatus = APFloat::opInvalidOp != |
| 2417 | a.convertToInteger(Result&: api, RM: APFloat::rmTowardZero, IsExact: &ignored); |
| 2418 | return api; |
| 2419 | }); |
| 2420 | } |
| 2421 | |
| 2422 | //===----------------------------------------------------------------------===// |
| 2423 | // IndexCastOp |
| 2424 | //===----------------------------------------------------------------------===// |
| 2425 | |
| 2426 | /// Return the bit-width of \p t for the purpose of index_cast width checks. |
| 2427 | /// For vector types use the element type; index maps to its internal storage |
| 2428 | /// width (64 on all current targets). |
| 2429 | static unsigned getIndexCastWidth(Type t) { |
| 2430 | if (auto intTy = dyn_cast<IntegerType>(Val: getElementTypeOrSelf(type: t))) |
| 2431 | return intTy.getWidth(); |
| 2432 | return IndexType::kInternalStorageBitWidth; |
| 2433 | } |
| 2434 | |
| 2435 | static bool areIndexCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2436 | if (!areValidCastInputsAndOutputs(inputs, outputs)) |
| 2437 | return false; |
| 2438 | |
| 2439 | auto srcType = getTypeIfLikeOrMemRef<IntegerType, IndexType>(type: inputs.front()); |
| 2440 | auto dstType = getTypeIfLikeOrMemRef<IntegerType, IndexType>(type: outputs.front()); |
| 2441 | if (!srcType || !dstType) |
| 2442 | return false; |
| 2443 | |
| 2444 | return (srcType.isIndex() && dstType.isSignlessInteger()) || |
| 2445 | (srcType.isSignlessInteger() && dstType.isIndex()); |
| 2446 | } |
| 2447 | |
| 2448 | bool arith::IndexCastOp::areCastCompatible(TypeRange inputs, |
| 2449 | TypeRange outputs) { |
| 2450 | return areIndexCastCompatible(inputs, outputs); |
| 2451 | } |
| 2452 | |
| 2453 | OpFoldResult arith::IndexCastOp::fold(FoldAdaptor adaptor) { |
| 2454 | // index_cast(constant) -> constant |
| 2455 | unsigned resultBitwidth = 64; // Default for index integer attributes. |
| 2456 | if (auto intTy = dyn_cast<IntegerType>(Val: getElementTypeOrSelf(type: getType()))) |
| 2457 | resultBitwidth = intTy.getWidth(); |
| 2458 | |
| 2459 | if (auto foldResult = constFoldCastOp<IntegerAttr, IntegerAttr>( |
| 2460 | operands: adaptor.getOperands(), resType: getType(), |
| 2461 | calculate: [resultBitwidth](const APInt &a, bool & /*castStatus*/) { |
| 2462 | return a.sextOrTrunc(width: resultBitwidth); |
| 2463 | })) |
| 2464 | return foldResult; |
| 2465 | |
| 2466 | // index_cast(index_cast(x : A) : B) : A -> x, but only when B is at least |
| 2467 | // as wide as A. If B is narrower, the inner cast truncates and the outer |
| 2468 | // cast sign-extends, so the round-trip is lossy. |
| 2469 | if (auto inner = getOperand().getDefiningOp<arith::IndexCastOp>()) { |
| 2470 | Value x = inner.getOperand(); |
| 2471 | if (x.getType() == getType()) { |
| 2472 | if (getIndexCastWidth(t: inner.getType()) >= getIndexCastWidth(t: x.getType())) |
| 2473 | return x; |
| 2474 | } |
| 2475 | } |
| 2476 | return {}; |
| 2477 | } |
| 2478 | |
| 2479 | void arith::IndexCastOp::getCanonicalizationPatterns( |
| 2480 | RewritePatternSet &patterns, MLIRContext *context) { |
| 2481 | patterns.add<IndexCastOfExtSI>(arg&: context); |
| 2482 | } |
| 2483 | |
| 2484 | //===----------------------------------------------------------------------===// |
| 2485 | // IndexCastUIOp |
| 2486 | //===----------------------------------------------------------------------===// |
| 2487 | |
| 2488 | bool arith::IndexCastUIOp::areCastCompatible(TypeRange inputs, |
| 2489 | TypeRange outputs) { |
| 2490 | return areIndexCastCompatible(inputs, outputs); |
| 2491 | } |
| 2492 | |
| 2493 | OpFoldResult arith::IndexCastUIOp::fold(FoldAdaptor adaptor) { |
| 2494 | // index_castui(constant) -> constant |
| 2495 | unsigned resultBitwidth = 64; // Default for index integer attributes. |
| 2496 | if (auto intTy = dyn_cast<IntegerType>(Val: getElementTypeOrSelf(type: getType()))) |
| 2497 | resultBitwidth = intTy.getWidth(); |
| 2498 | |
| 2499 | if (auto foldResult = constFoldCastOp<IntegerAttr, IntegerAttr>( |
| 2500 | operands: adaptor.getOperands(), resType: getType(), |
| 2501 | calculate: [resultBitwidth](const APInt &a, bool & /*castStatus*/) { |
| 2502 | return a.zextOrTrunc(width: resultBitwidth); |
| 2503 | })) |
| 2504 | return foldResult; |
| 2505 | |
| 2506 | // index_castui(index_castui(x : A) : B) : A -> x, but only when B is at |
| 2507 | // least as wide as A. If B is narrower, the inner cast truncates and the |
| 2508 | // outer cast zero-extends, so the round-trip is lossy. |
| 2509 | if (auto inner = getOperand().getDefiningOp<arith::IndexCastUIOp>()) { |
| 2510 | Value x = inner.getOperand(); |
| 2511 | if (x.getType() == getType()) { |
| 2512 | if (getIndexCastWidth(t: inner.getType()) >= getIndexCastWidth(t: x.getType())) |
| 2513 | return x; |
| 2514 | } |
| 2515 | } |
| 2516 | return {}; |
| 2517 | } |
| 2518 | |
| 2519 | void arith::IndexCastUIOp::getCanonicalizationPatterns( |
| 2520 | RewritePatternSet &patterns, MLIRContext *context) { |
| 2521 | patterns.add<IndexCastUIOfExtUI>(arg&: context); |
| 2522 | } |
| 2523 | |
| 2524 | //===----------------------------------------------------------------------===// |
| 2525 | // BitcastOp |
| 2526 | //===----------------------------------------------------------------------===// |
| 2527 | |
| 2528 | bool arith::BitcastOp::areCastCompatible(TypeRange inputs, TypeRange outputs) { |
| 2529 | if (!areValidCastInputsAndOutputs(inputs, outputs)) |
| 2530 | return false; |
| 2531 | |
| 2532 | auto srcType = getTypeIfLikeOrMemRef<IntegerType, FloatType>(type: inputs.front()); |
| 2533 | auto dstType = getTypeIfLikeOrMemRef<IntegerType, FloatType>(type: outputs.front()); |
| 2534 | if (!srcType || !dstType) |
| 2535 | return false; |
| 2536 | |
| 2537 | return srcType.getIntOrFloatBitWidth() == dstType.getIntOrFloatBitWidth(); |
| 2538 | } |
| 2539 | |
| 2540 | OpFoldResult arith::BitcastOp::fold(FoldAdaptor adaptor) { |
| 2541 | auto resType = getType(); |
| 2542 | auto operand = adaptor.getIn(); |
| 2543 | if (!operand) |
| 2544 | return {}; |
| 2545 | |
| 2546 | /// Bitcast dense elements. |
| 2547 | if (auto denseAttr = dyn_cast_or_null<DenseElementsAttr>(Val&: operand)) |
| 2548 | return denseAttr.bitcast(newElType: llvm::cast<ShapedType>(Val&: resType).getElementType()); |
| 2549 | /// Other shaped types unhandled. |
| 2550 | if (llvm::isa<ShapedType>(Val: resType)) |
| 2551 | return {}; |
| 2552 | |
| 2553 | /// Bitcast poison. |
| 2554 | if (matchPattern(attr: operand, pattern: ub::m_Poison())) |
| 2555 | return ub::PoisonAttr::get(ctx: getContext()); |
| 2556 | |
| 2557 | /// Bitcast integer or float to integer or float. |
| 2558 | if (!llvm::isa<FloatAttr, IntegerAttr>(Val: operand)) |
| 2559 | return {}; |
| 2560 | |
| 2561 | APInt bits = llvm::isa<FloatAttr>(Val: operand) |
| 2562 | ? llvm::cast<FloatAttr>(Val&: operand).getValue().bitcastToAPInt() |
| 2563 | : llvm::cast<IntegerAttr>(Val&: operand).getValue(); |
| 2564 | assert(resType.getIntOrFloatBitWidth() == bits.getBitWidth() && |
| 2565 | "trying to fold on broken IR: operands have incompatible types" ); |
| 2566 | |
| 2567 | if (auto resFloatType = dyn_cast<FloatType>(Val&: resType)) |
| 2568 | return FloatAttr::get(type: resType, |
| 2569 | value: APFloat(resFloatType.getFloatSemantics(), bits)); |
| 2570 | return IntegerAttr::get(type: resType, value: bits); |
| 2571 | } |
| 2572 | |
| 2573 | void arith::BitcastOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2574 | MLIRContext *context) { |
| 2575 | patterns.add<BitcastOfBitcast>(arg&: context); |
| 2576 | } |
| 2577 | |
| 2578 | //===----------------------------------------------------------------------===// |
| 2579 | // CmpIOp |
| 2580 | //===----------------------------------------------------------------------===// |
| 2581 | |
| 2582 | /// Compute `lhs` `pred` `rhs`, where `pred` is one of the known integer |
| 2583 | /// comparison predicates. |
| 2584 | bool mlir::arith::applyCmpPredicate(arith::CmpIPredicate predicate, |
| 2585 | const APInt &lhs, const APInt &rhs) { |
| 2586 | switch (predicate) { |
| 2587 | case arith::CmpIPredicate::eq: |
| 2588 | return lhs.eq(RHS: rhs); |
| 2589 | case arith::CmpIPredicate::ne: |
| 2590 | return lhs.ne(RHS: rhs); |
| 2591 | case arith::CmpIPredicate::slt: |
| 2592 | return lhs.slt(RHS: rhs); |
| 2593 | case arith::CmpIPredicate::sle: |
| 2594 | return lhs.sle(RHS: rhs); |
| 2595 | case arith::CmpIPredicate::sgt: |
| 2596 | return lhs.sgt(RHS: rhs); |
| 2597 | case arith::CmpIPredicate::sge: |
| 2598 | return lhs.sge(RHS: rhs); |
| 2599 | case arith::CmpIPredicate::ult: |
| 2600 | return lhs.ult(RHS: rhs); |
| 2601 | case arith::CmpIPredicate::ule: |
| 2602 | return lhs.ule(RHS: rhs); |
| 2603 | case arith::CmpIPredicate::ugt: |
| 2604 | return lhs.ugt(RHS: rhs); |
| 2605 | case arith::CmpIPredicate::uge: |
| 2606 | return lhs.uge(RHS: rhs); |
| 2607 | } |
| 2608 | llvm_unreachable("unknown cmpi predicate kind" ); |
| 2609 | } |
| 2610 | |
| 2611 | /// Returns true if the predicate is true for two equal operands. |
| 2612 | static bool applyCmpPredicateToEqualOperands(arith::CmpIPredicate predicate) { |
| 2613 | switch (predicate) { |
| 2614 | case arith::CmpIPredicate::eq: |
| 2615 | case arith::CmpIPredicate::sle: |
| 2616 | case arith::CmpIPredicate::sge: |
| 2617 | case arith::CmpIPredicate::ule: |
| 2618 | case arith::CmpIPredicate::uge: |
| 2619 | return true; |
| 2620 | case arith::CmpIPredicate::ne: |
| 2621 | case arith::CmpIPredicate::slt: |
| 2622 | case arith::CmpIPredicate::sgt: |
| 2623 | case arith::CmpIPredicate::ult: |
| 2624 | case arith::CmpIPredicate::ugt: |
| 2625 | return false; |
| 2626 | } |
| 2627 | llvm_unreachable("unknown cmpi predicate kind" ); |
| 2628 | } |
| 2629 | |
| 2630 | static std::optional<int64_t> getIntegerWidth(Type t) { |
| 2631 | if (auto intType = dyn_cast<IntegerType>(Val&: t)) { |
| 2632 | return intType.getWidth(); |
| 2633 | } |
| 2634 | if (auto vectorIntType = dyn_cast<VectorType>(Val&: t)) { |
| 2635 | return llvm::cast<IntegerType>(Val: vectorIntType.getElementType()).getWidth(); |
| 2636 | } |
| 2637 | return std::nullopt; |
| 2638 | } |
| 2639 | |
| 2640 | OpFoldResult arith::CmpIOp::fold(FoldAdaptor adaptor) { |
| 2641 | // cmpi(pred, x, x) |
| 2642 | if (getLhs() == getRhs()) { |
| 2643 | auto val = applyCmpPredicateToEqualOperands(predicate: getPredicate()); |
| 2644 | return getBoolAttribute(type: getType(), value: val); |
| 2645 | } |
| 2646 | |
| 2647 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) { |
| 2648 | if (auto extOp = getLhs().getDefiningOp<ExtSIOp>()) { |
| 2649 | // extsi(%x : i1 -> iN) != 0 -> %x |
| 2650 | std::optional<int64_t> integerWidth = |
| 2651 | getIntegerWidth(t: extOp.getOperand().getType()); |
| 2652 | if (integerWidth && integerWidth.value() == 1 && |
| 2653 | getPredicate() == arith::CmpIPredicate::ne) |
| 2654 | return extOp.getOperand(); |
| 2655 | } |
| 2656 | if (auto extOp = getLhs().getDefiningOp<ExtUIOp>()) { |
| 2657 | // extui(%x : i1 -> iN) != 0 -> %x |
| 2658 | std::optional<int64_t> integerWidth = |
| 2659 | getIntegerWidth(t: extOp.getOperand().getType()); |
| 2660 | if (integerWidth && integerWidth.value() == 1 && |
| 2661 | getPredicate() == arith::CmpIPredicate::ne) |
| 2662 | return extOp.getOperand(); |
| 2663 | } |
| 2664 | |
| 2665 | // arith.cmpi ne, %val, %zero : i1 -> %val |
| 2666 | if (getElementTypeOrSelf(type: getLhs().getType()).isInteger(width: 1) && |
| 2667 | getPredicate() == arith::CmpIPredicate::ne) |
| 2668 | return getLhs(); |
| 2669 | } |
| 2670 | |
| 2671 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_One())) { |
| 2672 | // arith.cmpi eq, %val, %one : i1 -> %val |
| 2673 | if (getElementTypeOrSelf(type: getLhs().getType()).isInteger(width: 1) && |
| 2674 | getPredicate() == arith::CmpIPredicate::eq) |
| 2675 | return getLhs(); |
| 2676 | } |
| 2677 | |
| 2678 | // Move constant to the right side. |
| 2679 | if (adaptor.getLhs() && !adaptor.getRhs()) { |
| 2680 | // Do not use invertPredicate, as it will change eq to ne and vice versa. |
| 2681 | using Pred = CmpIPredicate; |
| 2682 | const std::pair<Pred, Pred> invPreds[] = { |
| 2683 | {Pred::slt, Pred::sgt}, {Pred::sgt, Pred::slt}, {Pred::sle, Pred::sge}, |
| 2684 | {Pred::sge, Pred::sle}, {Pred::ult, Pred::ugt}, {Pred::ugt, Pred::ult}, |
| 2685 | {Pred::ule, Pred::uge}, {Pred::uge, Pred::ule}, {Pred::eq, Pred::eq}, |
| 2686 | {Pred::ne, Pred::ne}, |
| 2687 | }; |
| 2688 | Pred origPred = getPredicate(); |
| 2689 | for (auto pred : invPreds) { |
| 2690 | if (origPred == pred.first) { |
| 2691 | setPredicate(pred.second); |
| 2692 | Value lhs = getLhs(); |
| 2693 | Value rhs = getRhs(); |
| 2694 | getLhsMutable().assign(value: rhs); |
| 2695 | getRhsMutable().assign(value: lhs); |
| 2696 | return getResult(); |
| 2697 | } |
| 2698 | } |
| 2699 | llvm_unreachable("unknown cmpi predicate kind" ); |
| 2700 | } |
| 2701 | |
| 2702 | // We are moving constants to the right side; So if lhs is constant rhs is |
| 2703 | // guaranteed to be a constant. |
| 2704 | if (auto lhs = dyn_cast_if_present<TypedAttr>(Val: adaptor.getLhs())) { |
| 2705 | return constFoldBinaryOp<IntegerAttr>( |
| 2706 | operands: adaptor.getOperands(), resultType: getI1SameShape(type: lhs.getType()), |
| 2707 | calculate: [pred = getPredicate()](const APInt &lhs, const APInt &rhs) { |
| 2708 | return APInt(1, |
| 2709 | static_cast<int64_t>(applyCmpPredicate(predicate: pred, lhs, rhs))); |
| 2710 | }); |
| 2711 | } |
| 2712 | |
| 2713 | return {}; |
| 2714 | } |
| 2715 | |
| 2716 | void arith::CmpIOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 2717 | MLIRContext *context) { |
| 2718 | patterns.insert<CmpIExtSI, CmpIExtUI>(arg&: context); |
| 2719 | } |
| 2720 | |
| 2721 | //===----------------------------------------------------------------------===// |
| 2722 | // CmpFOp |
| 2723 | //===----------------------------------------------------------------------===// |
| 2724 | |
| 2725 | /// Compute `lhs` `pred` `rhs`, where `pred` is one of the known floating point |
| 2726 | /// comparison predicates. |
| 2727 | bool mlir::arith::applyCmpPredicate(arith::CmpFPredicate predicate, |
| 2728 | const APFloat &lhs, const APFloat &rhs) { |
| 2729 | auto cmpResult = lhs.compare(RHS: rhs); |
| 2730 | switch (predicate) { |
| 2731 | case arith::CmpFPredicate::AlwaysFalse: |
| 2732 | return false; |
| 2733 | case arith::CmpFPredicate::OEQ: |
| 2734 | return cmpResult == APFloat::cmpEqual; |
| 2735 | case arith::CmpFPredicate::OGT: |
| 2736 | return cmpResult == APFloat::cmpGreaterThan; |
| 2737 | case arith::CmpFPredicate::OGE: |
| 2738 | return cmpResult == APFloat::cmpGreaterThan || |
| 2739 | cmpResult == APFloat::cmpEqual; |
| 2740 | case arith::CmpFPredicate::OLT: |
| 2741 | return cmpResult == APFloat::cmpLessThan; |
| 2742 | case arith::CmpFPredicate::OLE: |
| 2743 | return cmpResult == APFloat::cmpLessThan || cmpResult == APFloat::cmpEqual; |
| 2744 | case arith::CmpFPredicate::ONE: |
| 2745 | return cmpResult != APFloat::cmpUnordered && cmpResult != APFloat::cmpEqual; |
| 2746 | case arith::CmpFPredicate::ORD: |
| 2747 | return cmpResult != APFloat::cmpUnordered; |
| 2748 | case arith::CmpFPredicate::UEQ: |
| 2749 | return cmpResult == APFloat::cmpUnordered || cmpResult == APFloat::cmpEqual; |
| 2750 | case arith::CmpFPredicate::UGT: |
| 2751 | return cmpResult == APFloat::cmpUnordered || |
| 2752 | cmpResult == APFloat::cmpGreaterThan; |
| 2753 | case arith::CmpFPredicate::UGE: |
| 2754 | return cmpResult == APFloat::cmpUnordered || |
| 2755 | cmpResult == APFloat::cmpGreaterThan || |
| 2756 | cmpResult == APFloat::cmpEqual; |
| 2757 | case arith::CmpFPredicate::ULT: |
| 2758 | return cmpResult == APFloat::cmpUnordered || |
| 2759 | cmpResult == APFloat::cmpLessThan; |
| 2760 | case arith::CmpFPredicate::ULE: |
| 2761 | return cmpResult == APFloat::cmpUnordered || |
| 2762 | cmpResult == APFloat::cmpLessThan || cmpResult == APFloat::cmpEqual; |
| 2763 | case arith::CmpFPredicate::UNE: |
| 2764 | return cmpResult != APFloat::cmpEqual; |
| 2765 | case arith::CmpFPredicate::UNO: |
| 2766 | return cmpResult == APFloat::cmpUnordered; |
| 2767 | case arith::CmpFPredicate::AlwaysTrue: |
| 2768 | return true; |
| 2769 | } |
| 2770 | llvm_unreachable("unknown cmpf predicate kind" ); |
| 2771 | } |
| 2772 | |
| 2773 | OpFoldResult arith::CmpFOp::fold(FoldAdaptor adaptor) { |
| 2774 | auto lhs = dyn_cast_if_present<FloatAttr>(Val: adaptor.getLhs()); |
| 2775 | auto rhs = dyn_cast_if_present<FloatAttr>(Val: adaptor.getRhs()); |
| 2776 | |
| 2777 | // If one operand is NaN, making them both NaN does not change the result. |
| 2778 | if (lhs && lhs.getValue().isNaN()) |
| 2779 | rhs = lhs; |
| 2780 | if (rhs && rhs.getValue().isNaN()) |
| 2781 | lhs = rhs; |
| 2782 | |
| 2783 | if (!lhs || !rhs) |
| 2784 | return {}; |
| 2785 | |
| 2786 | auto val = applyCmpPredicate(predicate: getPredicate(), lhs: lhs.getValue(), rhs: rhs.getValue()); |
| 2787 | return BoolAttr::get(context: getContext(), value: val); |
| 2788 | } |
| 2789 | |
| 2790 | namespace { |
| 2791 | class CmpFIntToFPConst final : public OpRewritePattern<CmpFOp> { |
| 2792 | public: |
| 2793 | using Base::Base; |
| 2794 | |
| 2795 | static CmpIPredicate convertToIntegerPredicate(CmpFPredicate pred, |
| 2796 | bool isUnsigned) { |
| 2797 | using namespace arith; |
| 2798 | switch (pred) { |
| 2799 | case CmpFPredicate::UEQ: |
| 2800 | case CmpFPredicate::OEQ: |
| 2801 | return CmpIPredicate::eq; |
| 2802 | case CmpFPredicate::UGT: |
| 2803 | case CmpFPredicate::OGT: |
| 2804 | return isUnsigned ? CmpIPredicate::ugt : CmpIPredicate::sgt; |
| 2805 | case CmpFPredicate::UGE: |
| 2806 | case CmpFPredicate::OGE: |
| 2807 | return isUnsigned ? CmpIPredicate::uge : CmpIPredicate::sge; |
| 2808 | case CmpFPredicate::ULT: |
| 2809 | case CmpFPredicate::OLT: |
| 2810 | return isUnsigned ? CmpIPredicate::ult : CmpIPredicate::slt; |
| 2811 | case CmpFPredicate::ULE: |
| 2812 | case CmpFPredicate::OLE: |
| 2813 | return isUnsigned ? CmpIPredicate::ule : CmpIPredicate::sle; |
| 2814 | case CmpFPredicate::UNE: |
| 2815 | case CmpFPredicate::ONE: |
| 2816 | return CmpIPredicate::ne; |
| 2817 | default: |
| 2818 | llvm_unreachable("Unexpected predicate!" ); |
| 2819 | } |
| 2820 | } |
| 2821 | |
| 2822 | LogicalResult matchAndRewrite(CmpFOp op, |
| 2823 | PatternRewriter &rewriter) const override { |
| 2824 | FloatAttr flt; |
| 2825 | if (!matchPattern(value: op.getRhs(), pattern: m_Constant(bind_value: &flt))) |
| 2826 | return failure(); |
| 2827 | |
| 2828 | const APFloat &rhs = flt.getValue(); |
| 2829 | |
| 2830 | // Don't attempt to fold a nan. |
| 2831 | if (rhs.isNaN()) |
| 2832 | return failure(); |
| 2833 | |
| 2834 | // Get the width of the mantissa. We don't want to hack on conversions that |
| 2835 | // might lose information from the integer, e.g. "i64 -> float" |
| 2836 | FloatType floatTy = llvm::cast<FloatType>(Val: op.getRhs().getType()); |
| 2837 | int mantissaWidth = floatTy.getFPMantissaWidth(); |
| 2838 | if (mantissaWidth <= 0) |
| 2839 | return failure(); |
| 2840 | |
| 2841 | bool isUnsigned; |
| 2842 | Value intVal; |
| 2843 | |
| 2844 | if (auto si = op.getLhs().getDefiningOp<SIToFPOp>()) { |
| 2845 | isUnsigned = false; |
| 2846 | intVal = si.getIn(); |
| 2847 | } else if (auto ui = op.getLhs().getDefiningOp<UIToFPOp>()) { |
| 2848 | isUnsigned = true; |
| 2849 | intVal = ui.getIn(); |
| 2850 | } else { |
| 2851 | return failure(); |
| 2852 | } |
| 2853 | |
| 2854 | // Check to see that the input is converted from an integer type that is |
| 2855 | // small enough that preserves all bits. |
| 2856 | auto intTy = llvm::cast<IntegerType>(Val: intVal.getType()); |
| 2857 | auto intWidth = intTy.getWidth(); |
| 2858 | |
| 2859 | // Number of bits representing values, as opposed to the sign |
| 2860 | auto valueBits = isUnsigned ? intWidth : (intWidth - 1); |
| 2861 | |
| 2862 | const llvm::fltSemantics &semantics = floatTy.getFloatSemantics(); |
| 2863 | |
| 2864 | // Conservatively require an exact representation of integer zero. For |
| 2865 | // example, f8E8M0FNU converts zero to its smallest positive value. |
| 2866 | if (!APFloat::semanticsHasZero(semantics)) |
| 2867 | return failure(); |
| 2868 | |
| 2869 | // Signed inputs may be negative, which e.g. f8E5M3FNU cannot represent. |
| 2870 | // Do not try to convert the signed minimum below for such formats. |
| 2871 | if (!isUnsigned && !APFloat::semanticsHasSignedRepr(semantics)) |
| 2872 | return failure(); |
| 2873 | |
| 2874 | if (!APFloat::semanticsHasInf(semantics)) { |
| 2875 | // Overflow can produce NaN or saturate to a finite value: |
| 2876 | // - sitofp(1000 : i32) to f8E4M3FN is NaN, so cmpf ogt against 0.0 |
| 2877 | // is false, whereas cmpi sgt against integer zero would be true. |
| 2878 | // - uitofp(8 : i4) to f6E2M3FN saturates to 7.5. Equality against 7.5 |
| 2879 | // is true, but the fractional-constant logic below would fold it to |
| 2880 | // false. Its four bits of precision do not prevent overflow. |
| 2881 | // |
| 2882 | // Conservatively require the enclosing 2^valueBits (and its negative |
| 2883 | // for signed inputs) to be representable, so rounding stays within |
| 2884 | // finite bounds. Equality at the exponent bound is safe. |
| 2885 | if (APFloat::semanticsMaxExponent(semantics) < |
| 2886 | static_cast<int>(valueBits)) |
| 2887 | return failure(); |
| 2888 | } |
| 2889 | |
| 2890 | // Following test does NOT adjust intWidth downwards for signed inputs, |
| 2891 | // because the most negative value still requires all the mantissa bits |
| 2892 | // to distinguish it from one less than that value. |
| 2893 | if ((int)intWidth > mantissaWidth) { |
| 2894 | // Conversion would lose accuracy. Check if loss can impact comparison. |
| 2895 | int exponent = ilogb(Arg: rhs); |
| 2896 | if (exponent == APFloat::IEK_Inf) { |
| 2897 | int maxExponent = ilogb(Arg: APFloat::getLargest(Sem: rhs.getSemantics())); |
| 2898 | if (maxExponent < static_cast<int>(valueBits)) { |
| 2899 | // Conversion could create infinity. For example, sitofp(65536 : i32) |
| 2900 | // to f16 is +inf, so equality against +inf can be true. |
| 2901 | return failure(); |
| 2902 | } |
| 2903 | } else { |
| 2904 | // Note that if rhs is zero or NaN, then Exp is negative |
| 2905 | // and first condition is trivially false. |
| 2906 | if (mantissaWidth <= exponent && exponent <= (int)valueBits) { |
| 2907 | // Conversion could affect comparison. |
| 2908 | return failure(); |
| 2909 | } |
| 2910 | } |
| 2911 | } |
| 2912 | |
| 2913 | // Convert to equivalent cmpi predicate |
| 2914 | CmpIPredicate pred; |
| 2915 | switch (op.getPredicate()) { |
| 2916 | case CmpFPredicate::AlwaysTrue: |
| 2917 | case CmpFPredicate::ORD: |
| 2918 | // Int to fp conversion doesn't create a nan (ord checks neither is a nan) |
| 2919 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 2920 | /*width=*/args: 1); |
| 2921 | return success(); |
| 2922 | case CmpFPredicate::AlwaysFalse: |
| 2923 | case CmpFPredicate::UNO: |
| 2924 | // Int to fp conversion doesn't create a nan (uno checks either is a nan) |
| 2925 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 2926 | /*width=*/args: 1); |
| 2927 | return success(); |
| 2928 | default: |
| 2929 | pred = convertToIntegerPredicate(pred: op.getPredicate(), isUnsigned); |
| 2930 | break; |
| 2931 | } |
| 2932 | |
| 2933 | if (!isUnsigned) { |
| 2934 | // If the rhs value is > SignedMax, fold the comparison. This handles |
| 2935 | // +INF and large values. |
| 2936 | APFloat signedMax(rhs.getSemantics()); |
| 2937 | signedMax.convertFromAPInt(Input: APInt::getSignedMaxValue(numBits: intWidth), IsSigned: true, |
| 2938 | RM: APFloat::rmNearestTiesToEven); |
| 2939 | if (signedMax < rhs) { // smax < 13123.0 |
| 2940 | if (pred == CmpIPredicate::ne || pred == CmpIPredicate::slt || |
| 2941 | pred == CmpIPredicate::sle) |
| 2942 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 2943 | /*width=*/args: 1); |
| 2944 | else |
| 2945 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 2946 | /*width=*/args: 1); |
| 2947 | return success(); |
| 2948 | } |
| 2949 | } else { |
| 2950 | // If the rhs value is > UnsignedMax, fold the comparison. This handles |
| 2951 | // +INF and large values. |
| 2952 | APFloat unsignedMax(rhs.getSemantics()); |
| 2953 | unsignedMax.convertFromAPInt(Input: APInt::getMaxValue(numBits: intWidth), IsSigned: false, |
| 2954 | RM: APFloat::rmNearestTiesToEven); |
| 2955 | if (unsignedMax < rhs) { // umax < 13123.0 |
| 2956 | if (pred == CmpIPredicate::ne || pred == CmpIPredicate::ult || |
| 2957 | pred == CmpIPredicate::ule) |
| 2958 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 2959 | /*width=*/args: 1); |
| 2960 | else |
| 2961 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 2962 | /*width=*/args: 1); |
| 2963 | return success(); |
| 2964 | } |
| 2965 | } |
| 2966 | |
| 2967 | if (!isUnsigned) { |
| 2968 | // See if the rhs value is < SignedMin. |
| 2969 | APFloat signedMin(rhs.getSemantics()); |
| 2970 | signedMin.convertFromAPInt(Input: APInt::getSignedMinValue(numBits: intWidth), IsSigned: true, |
| 2971 | RM: APFloat::rmNearestTiesToEven); |
| 2972 | if (signedMin > rhs) { // smin > 12312.0 |
| 2973 | if (pred == CmpIPredicate::ne || pred == CmpIPredicate::sgt || |
| 2974 | pred == CmpIPredicate::sge) |
| 2975 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 2976 | /*width=*/args: 1); |
| 2977 | else |
| 2978 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 2979 | /*width=*/args: 1); |
| 2980 | return success(); |
| 2981 | } |
| 2982 | } else { |
| 2983 | // See if the rhs value is < UnsignedMin. |
| 2984 | APFloat unsignedMin(rhs.getSemantics()); |
| 2985 | unsignedMin.convertFromAPInt(Input: APInt::getMinValue(numBits: intWidth), IsSigned: false, |
| 2986 | RM: APFloat::rmNearestTiesToEven); |
| 2987 | if (unsignedMin > rhs) { // umin > 12312.0 |
| 2988 | if (pred == CmpIPredicate::ne || pred == CmpIPredicate::ugt || |
| 2989 | pred == CmpIPredicate::uge) |
| 2990 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 2991 | /*width=*/args: 1); |
| 2992 | else |
| 2993 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 2994 | /*width=*/args: 1); |
| 2995 | return success(); |
| 2996 | } |
| 2997 | } |
| 2998 | |
| 2999 | // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or |
| 3000 | // [0, UMAX], but it may still be fractional. See if it is fractional by |
| 3001 | // casting the FP value to the integer value and back, checking for |
| 3002 | // equality. Don't do this for zero, because -0.0 is not fractional. |
| 3003 | bool ignored; |
| 3004 | APSInt rhsInt(intWidth, isUnsigned); |
| 3005 | if (APFloat::opInvalidOp == |
| 3006 | rhs.convertToInteger(Result&: rhsInt, RM: APFloat::rmTowardZero, IsExact: &ignored)) { |
| 3007 | // Undefined behavior invoked - the destination type can't represent |
| 3008 | // the input constant. |
| 3009 | return failure(); |
| 3010 | } |
| 3011 | |
| 3012 | if (!rhs.isZero()) { |
| 3013 | APFloat apf(floatTy.getFloatSemantics(), |
| 3014 | APInt::getZero(numBits: floatTy.getWidth())); |
| 3015 | apf.convertFromAPInt(Input: rhsInt, IsSigned: !isUnsigned, RM: APFloat::rmNearestTiesToEven); |
| 3016 | |
| 3017 | bool equal = apf == rhs; |
| 3018 | if (!equal) { |
| 3019 | // If we had a comparison against a fractional value, we have to adjust |
| 3020 | // the compare predicate and sometimes the value. rhsInt is rounded |
| 3021 | // towards zero at this point. |
| 3022 | switch (pred) { |
| 3023 | case CmpIPredicate::ne: // (float)int != 4.4 --> true |
| 3024 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 3025 | /*width=*/args: 1); |
| 3026 | return success(); |
| 3027 | case CmpIPredicate::eq: // (float)int == 4.4 --> false |
| 3028 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 3029 | /*width=*/args: 1); |
| 3030 | return success(); |
| 3031 | case CmpIPredicate::ule: |
| 3032 | // (float)int <= 4.4 --> int <= 4 |
| 3033 | // (float)int <= -4.4 --> false |
| 3034 | if (rhs.isNegative()) { |
| 3035 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 3036 | /*width=*/args: 1); |
| 3037 | return success(); |
| 3038 | } |
| 3039 | break; |
| 3040 | case CmpIPredicate::sle: |
| 3041 | // (float)int <= 4.4 --> int <= 4 |
| 3042 | // (float)int <= -4.4 --> int < -4 |
| 3043 | if (rhs.isNegative()) |
| 3044 | pred = CmpIPredicate::slt; |
| 3045 | break; |
| 3046 | case CmpIPredicate::ult: |
| 3047 | // (float)int < -4.4 --> false |
| 3048 | // (float)int < 4.4 --> int <= 4 |
| 3049 | if (rhs.isNegative()) { |
| 3050 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: false, |
| 3051 | /*width=*/args: 1); |
| 3052 | return success(); |
| 3053 | } |
| 3054 | pred = CmpIPredicate::ule; |
| 3055 | break; |
| 3056 | case CmpIPredicate::slt: |
| 3057 | // (float)int < -4.4 --> int < -4 |
| 3058 | // (float)int < 4.4 --> int <= 4 |
| 3059 | if (!rhs.isNegative()) |
| 3060 | pred = CmpIPredicate::sle; |
| 3061 | break; |
| 3062 | case CmpIPredicate::ugt: |
| 3063 | // (float)int > 4.4 --> int > 4 |
| 3064 | // (float)int > -4.4 --> true |
| 3065 | if (rhs.isNegative()) { |
| 3066 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 3067 | /*width=*/args: 1); |
| 3068 | return success(); |
| 3069 | } |
| 3070 | break; |
| 3071 | case CmpIPredicate::sgt: |
| 3072 | // (float)int > 4.4 --> int > 4 |
| 3073 | // (float)int > -4.4 --> int >= -4 |
| 3074 | if (rhs.isNegative()) |
| 3075 | pred = CmpIPredicate::sge; |
| 3076 | break; |
| 3077 | case CmpIPredicate::uge: |
| 3078 | // (float)int >= -4.4 --> true |
| 3079 | // (float)int >= 4.4 --> int > 4 |
| 3080 | if (rhs.isNegative()) { |
| 3081 | rewriter.replaceOpWithNewOp<ConstantIntOp>(op, /*value=*/args: true, |
| 3082 | /*width=*/args: 1); |
| 3083 | return success(); |
| 3084 | } |
| 3085 | pred = CmpIPredicate::ugt; |
| 3086 | break; |
| 3087 | case CmpIPredicate::sge: |
| 3088 | // (float)int >= -4.4 --> int >= -4 |
| 3089 | // (float)int >= 4.4 --> int > 4 |
| 3090 | if (!rhs.isNegative()) |
| 3091 | pred = CmpIPredicate::sgt; |
| 3092 | break; |
| 3093 | } |
| 3094 | } |
| 3095 | } |
| 3096 | |
| 3097 | // Lower this FP comparison into an appropriate integer version of the |
| 3098 | // comparison. |
| 3099 | rewriter.replaceOpWithNewOp<CmpIOp>( |
| 3100 | op, args&: pred, args&: intVal, |
| 3101 | args: ConstantOp::create(builder&: rewriter, location: op.getLoc(), result: intVal.getType(), |
| 3102 | value: rewriter.getIntegerAttr(type: intVal.getType(), value: rhsInt))); |
| 3103 | return success(); |
| 3104 | } |
| 3105 | }; |
| 3106 | } // namespace |
| 3107 | |
| 3108 | void arith::CmpFOp::getCanonicalizationPatterns(RewritePatternSet &patterns, |
| 3109 | MLIRContext *context) { |
| 3110 | patterns.insert<CmpFIntToFPConst>(arg&: context); |
| 3111 | } |
| 3112 | |
| 3113 | //===----------------------------------------------------------------------===// |
| 3114 | // SelectOp |
| 3115 | //===----------------------------------------------------------------------===// |
| 3116 | |
| 3117 | namespace { |
| 3118 | // select %arg, %c1, %c0 => extui %arg |
| 3119 | struct SelectToExtUI : public OpRewritePattern<arith::SelectOp> { |
| 3120 | using Base::Base; |
| 3121 | |
| 3122 | LogicalResult matchAndRewrite(arith::SelectOp op, |
| 3123 | PatternRewriter &rewriter) const override { |
| 3124 | // Cannot extui i1 to i1, or i1 to f32 |
| 3125 | if (!llvm::isa<IntegerType>(Val: op.getType()) || op.getType().isInteger(width: 1)) |
| 3126 | return failure(); |
| 3127 | |
| 3128 | // select %x, c1, %c0 => extui %arg |
| 3129 | if (matchPattern(value: op.getTrueValue(), pattern: m_One()) && |
| 3130 | matchPattern(value: op.getFalseValue(), pattern: m_Zero())) { |
| 3131 | rewriter.replaceOpWithNewOp<arith::ExtUIOp>(op, args: op.getType(), |
| 3132 | args: op.getCondition()); |
| 3133 | return success(); |
| 3134 | } |
| 3135 | |
| 3136 | // select %x, c0, %c1 => extui (xor %arg, true) |
| 3137 | if (matchPattern(value: op.getTrueValue(), pattern: m_Zero()) && |
| 3138 | matchPattern(value: op.getFalseValue(), pattern: m_One())) { |
| 3139 | rewriter.replaceOpWithNewOp<arith::ExtUIOp>( |
| 3140 | op, args: op.getType(), |
| 3141 | args: arith::XOrIOp::create( |
| 3142 | builder&: rewriter, location: op.getLoc(), lhs: op.getCondition(), |
| 3143 | rhs: arith::ConstantIntOp::create(builder&: rewriter, location: op.getLoc(), |
| 3144 | type: op.getCondition().getType(), value: 1))); |
| 3145 | return success(); |
| 3146 | } |
| 3147 | |
| 3148 | return failure(); |
| 3149 | } |
| 3150 | }; |
| 3151 | } // namespace |
| 3152 | |
| 3153 | /// select(M, F(a0, a1, ..., a_n), F(b0, b1, ..., b_n)) |
| 3154 | /// -> F(select(M, a0, b0), select(M, a1, b1), ...) for pure and |
| 3155 | /// elementwise F. |
| 3156 | /// |
| 3157 | /// Correctness provement: |
| 3158 | /// |
| 3159 | /// Let's discuss the cases ignoring poison first. |
| 3160 | /// |
| 3161 | /// As F is elementwise, we have |
| 3162 | /// |
| 3163 | /// F(a1, a2, ..., a_n)[p] = F(a1[p], a2[p], ..., a_n[p]) |
| 3164 | /// |
| 3165 | /// then for |
| 3166 | /// |
| 3167 | /// L = select(M, F(a0, a1, ..., a_n), F(b0, b1, ..., b_n)) |
| 3168 | /// R = F(select(M, a0, b0), select(M, a1, b1), ....) |
| 3169 | /// |
| 3170 | /// for any position p, we have: |
| 3171 | /// |
| 3172 | /// if M[p] == true, then L[p] == F(a)[p] == F(a[p]) == F(select(true, a[p], b[p])) == R[p] |
| 3173 | /// if M[p] == false, then L[p] == F(b)[p] == F(b[p]) == F(select(false, a[p], b[p])) == R[p] |
| 3174 | /// |
| 3175 | /// And F is pure, so that we can move and reduce F. |
| 3176 | /// So select(M, F(a0, a1, ..., a_n), F(b0, b1, ..., b_n)) is equal to |
| 3177 | /// F(select(M, a0, b0), select(M, a1, b1), ...) ignoring poison. |
| 3178 | /// |
| 3179 | /// Let's consider the poison value now. ArithBase.td says: |
| 3180 | /// |
| 3181 | /// Unless otherwise stated, the operations within this dialect |
| 3182 | /// propagate poison values, i.e., if any of its inputs are poison, then the |
| 3183 | /// output is poison. |
| 3184 | /// |
| 3185 | /// So most F, F(poison) = poison |
| 3186 | /// |
| 3187 | /// Then for any position p, if a[p] is poison |
| 3188 | /// |
| 3189 | /// if M[p] == true, L[p] == f(a[p]) == poison == F(select(true, a[p], b[p])) == R[p] |
| 3190 | /// if M[p] == false, L[p] == f(b[p]) == F(select(false, a[p], b[p])) == R[p] |
| 3191 | /// if M[p] == P, L[p] == poison == F(select(P, a, b)) == R[p] |
| 3192 | /// |
| 3193 | /// Then the transformation doesn't change poison propagation. |
| 3194 | /// |
| 3195 | /// For F which doesn't meet the assumption, this transformation may change the |
| 3196 | /// poison propagation, but the language allows us to replace poison with a undef |
| 3197 | /// or any value with the same type. So this is still valid transformation. |
| 3198 | struct SelectCommonElementwiseOp : public OpRewritePattern<arith::SelectOp> { |
| 3199 | using Base::Base; |
| 3200 | |
| 3201 | LogicalResult matchAndRewrite(arith::SelectOp op, |
| 3202 | PatternRewriter &rewriter) const override { |
| 3203 | Operation *lhs = op.getTrueValue().getDefiningOp(); |
| 3204 | Operation *rhs = op.getFalseValue().getDefiningOp(); |
| 3205 | if (!lhs || !rhs || lhs == rhs) |
| 3206 | return failure(); |
| 3207 | |
| 3208 | // Do not hoist a select through selects: rewriting |
| 3209 | // select(c, select(d, a, b), select(d, x, y)) into |
| 3210 | // select(d, select(c, a, x), select(c, b, y)) matches again with c and d |
| 3211 | // swapped, so the greedy rewriter never converges. |
| 3212 | if (isa<arith::SelectOp>(Val: lhs)) |
| 3213 | return failure(); |
| 3214 | |
| 3215 | for (Operation *producer : {lhs, rhs}) { |
| 3216 | // Avoid sinking loop invariant code into the loop. |
| 3217 | // |
| 3218 | // TODO: handle the cases by introducing more CFG/Loop analysis. |
| 3219 | if (producer->getBlock() != op->getBlock()) |
| 3220 | return failure(); |
| 3221 | |
| 3222 | // `isEquivalentTo` doesn't catch all cases, so we conservatively match |
| 3223 | // the cases we know how to convert |
| 3224 | if (producer->getNumResults() != 1 || !producer->hasOneUse() || |
| 3225 | producer->getNumRegions() != 0 || producer->getNumSuccessors() != 0 || |
| 3226 | !producer->hasTrait<OpTrait::Elementwise>() || !isPure(op: producer)) |
| 3227 | return failure(); |
| 3228 | } |
| 3229 | |
| 3230 | // Check if the lhs and the rhs are the same kind of op, have the |
| 3231 | // same structure of inputs, types, and attributes. |
| 3232 | if (!OperationEquivalence::isEquivalentTo( |
| 3233 | lhs, rhs, checkEquivalent: OperationEquivalence::ignoreValueEquivalence, markEquivalent: nullptr, |
| 3234 | flags: OperationEquivalence::IgnoreLocations)) |
| 3235 | return failure(); |
| 3236 | |
| 3237 | DenseMap<std::pair<Value, Value>, Value> selected; |
| 3238 | SmallVector<Operation *> createdSelects; |
| 3239 | SmallVector<Value> operands; |
| 3240 | Type conditionType = op.getCondition().getType(); |
| 3241 | for (auto [a, b] : llvm::zip(t: lhs->getOperands(), u: rhs->getOperands())) { |
| 3242 | // For select(c, F(a, k), F(b, k)), we should transform it into |
| 3243 | // F(select(c, a, b), k) instead of F(select(c, a, b), select(c, k, k)) |
| 3244 | if (a == b) { |
| 3245 | operands.push_back(Elt: a); |
| 3246 | continue; |
| 3247 | } |
| 3248 | if (!conditionType.isSignlessInteger(width: 1) && |
| 3249 | (!isa<VectorType, TensorType>(Val: a.getType()) || |
| 3250 | cast<ShapedType>(Val: a.getType()).clone(elementType: rewriter.getI1Type()) != |
| 3251 | conditionType)) { |
| 3252 | |
| 3253 | for (Operation *created : llvm::reverse(C&: createdSelects)) |
| 3254 | rewriter.eraseOp(op: created); |
| 3255 | return failure(); |
| 3256 | } |
| 3257 | |
| 3258 | // Always creating a new pair would not merging equivalent results. |
| 3259 | auto [it, inserted] = selected.try_emplace(Key: std::make_pair(x&: a, y&: b)); |
| 3260 | if (inserted) { |
| 3261 | auto select = arith::SelectOp::create(builder&: rewriter, location: op.getLoc(), |
| 3262 | condition: op.getCondition(), true_value: a, false_value: b); |
| 3263 | it->second = select; |
| 3264 | createdSelects.push_back(Elt: select); |
| 3265 | } |
| 3266 | operands.push_back(Elt: it->second); |
| 3267 | } |
| 3268 | |
| 3269 | Operation *common = rewriter.clone(op&: *lhs); |
| 3270 | common->setOperands(operands); |
| 3271 | rewriter.replaceOp(op, newValues: common->getResults()); |
| 3272 | return success(); |
| 3273 | } |
| 3274 | }; |
| 3275 | |
| 3276 | void arith::SelectOp::getCanonicalizationPatterns(RewritePatternSet &results, |
| 3277 | MLIRContext *context) { |
| 3278 | results.add<RedundantSelectFalse, RedundantSelectTrue, SelectNotCond, |
| 3279 | SelectI1ToNot, SelectCmpISgeToMaxSI, SelectCmpISgeToMinSI, |
| 3280 | SelectCmpISgtToMaxSI, SelectCmpISgtToMinSI, SelectCmpISleToMaxSI, |
| 3281 | SelectCmpISleToMinSI, SelectCmpISltToMaxSI, SelectCmpISltToMinSI, |
| 3282 | SelectCmpIUgeToMaxUI, SelectCmpIUgeToMinUI, SelectCmpIUgtToMaxUI, |
| 3283 | SelectCmpIUgtToMinUI, SelectCmpIUleToMaxUI, SelectCmpIUleToMinUI, |
| 3284 | SelectCmpIUltToMaxUI, SelectCmpIUltToMinUI, SelectToExtUI>( |
| 3285 | arg&: context); |
| 3286 | results.add<SelectCommonElementwiseOp>(arg&: context); |
| 3287 | } |
| 3288 | |
| 3289 | OpFoldResult arith::SelectOp::fold(FoldAdaptor adaptor) { |
| 3290 | Value trueVal = getTrueValue(); |
| 3291 | Value falseVal = getFalseValue(); |
| 3292 | if (trueVal == falseVal) |
| 3293 | return trueVal; |
| 3294 | |
| 3295 | Value condition = getCondition(); |
| 3296 | |
| 3297 | // select true, %0, %1 => %0 |
| 3298 | if (matchPattern(attr: adaptor.getCondition(), pattern: m_One())) |
| 3299 | return trueVal; |
| 3300 | |
| 3301 | // select false, %0, %1 => %1 |
| 3302 | if (matchPattern(attr: adaptor.getCondition(), pattern: m_Zero())) |
| 3303 | return falseVal; |
| 3304 | |
| 3305 | // If either operand is fully poisoned, return the other. |
| 3306 | if (matchPattern(attr: adaptor.getTrueValue(), pattern: ub::m_Poison())) |
| 3307 | return falseVal; |
| 3308 | |
| 3309 | if (matchPattern(attr: adaptor.getFalseValue(), pattern: ub::m_Poison())) |
| 3310 | return trueVal; |
| 3311 | |
| 3312 | // select %x, true, false => %x |
| 3313 | if (getType().isSignlessInteger(width: 1) && |
| 3314 | matchPattern(attr: adaptor.getTrueValue(), pattern: m_One()) && |
| 3315 | matchPattern(attr: adaptor.getFalseValue(), pattern: m_Zero())) |
| 3316 | return condition; |
| 3317 | |
| 3318 | if (auto cmp = condition.getDefiningOp<arith::CmpIOp>()) { |
| 3319 | auto pred = cmp.getPredicate(); |
| 3320 | if (pred == arith::CmpIPredicate::eq || pred == arith::CmpIPredicate::ne) { |
| 3321 | auto cmpLhs = cmp.getLhs(); |
| 3322 | auto cmpRhs = cmp.getRhs(); |
| 3323 | |
| 3324 | // %0 = arith.cmpi eq, %arg0, %arg1 |
| 3325 | // %1 = arith.select %0, %arg0, %arg1 => %arg1 |
| 3326 | |
| 3327 | // %0 = arith.cmpi ne, %arg0, %arg1 |
| 3328 | // %1 = arith.select %0, %arg0, %arg1 => %arg0 |
| 3329 | |
| 3330 | if ((cmpLhs == trueVal && cmpRhs == falseVal) || |
| 3331 | (cmpRhs == trueVal && cmpLhs == falseVal)) |
| 3332 | return pred == arith::CmpIPredicate::ne ? trueVal : falseVal; |
| 3333 | } |
| 3334 | } |
| 3335 | |
| 3336 | // Constant-fold constant operands over non-splat constant condition. |
| 3337 | // select %cst_vec, %cst0, %cst1 => %cst2 |
| 3338 | if (auto cond = |
| 3339 | dyn_cast_if_present<DenseElementsAttr>(Val: adaptor.getCondition())) { |
| 3340 | // DenseElementsAttr by construction always has a static shape. |
| 3341 | assert(cond.getType().hasStaticShape() && |
| 3342 | "DenseElementsAttr must have static shape" ); |
| 3343 | if (auto lhs = |
| 3344 | dyn_cast_if_present<DenseElementsAttr>(Val: adaptor.getTrueValue())) { |
| 3345 | if (auto rhs = |
| 3346 | dyn_cast_if_present<DenseElementsAttr>(Val: adaptor.getFalseValue())) { |
| 3347 | SmallVector<Attribute> results; |
| 3348 | results.reserve(N: static_cast<size_t>(cond.getNumElements())); |
| 3349 | auto condVals = llvm::make_range(x: cond.value_begin<BoolAttr>(), |
| 3350 | y: cond.value_end<BoolAttr>()); |
| 3351 | auto lhsVals = llvm::make_range(x: lhs.value_begin<Attribute>(), |
| 3352 | y: lhs.value_end<Attribute>()); |
| 3353 | auto rhsVals = llvm::make_range(x: rhs.value_begin<Attribute>(), |
| 3354 | y: rhs.value_end<Attribute>()); |
| 3355 | |
| 3356 | for (auto [condVal, lhsVal, rhsVal] : |
| 3357 | llvm::zip_equal(t&: condVals, u&: lhsVals, args&: rhsVals)) |
| 3358 | results.push_back(Elt: condVal.getValue() ? lhsVal : rhsVal); |
| 3359 | |
| 3360 | return DenseElementsAttr::get(type: lhs.getType(), values: results); |
| 3361 | } |
| 3362 | } |
| 3363 | } |
| 3364 | |
| 3365 | return nullptr; |
| 3366 | } |
| 3367 | |
| 3368 | ParseResult SelectOp::parse(OpAsmParser &parser, OperationState &result) { |
| 3369 | Type conditionType, resultType; |
| 3370 | SmallVector<OpAsmParser::UnresolvedOperand, 3> operands; |
| 3371 | if (parser.parseOperandList(result&: operands, /*requiredOperandCount=*/3) || |
| 3372 | parser.parseOptionalAttrDict(result&: result.attributes) || |
| 3373 | parser.parseColonType(result&: resultType)) |
| 3374 | return failure(); |
| 3375 | |
| 3376 | // Check for the explicit condition type if this is a masked tensor or vector. |
| 3377 | if (succeeded(Result: parser.parseOptionalComma())) { |
| 3378 | conditionType = resultType; |
| 3379 | if (parser.parseType(result&: resultType)) |
| 3380 | return failure(); |
| 3381 | } else { |
| 3382 | conditionType = parser.getBuilder().getI1Type(); |
| 3383 | } |
| 3384 | |
| 3385 | result.addTypes(newTypes: resultType); |
| 3386 | return parser.resolveOperands(operands, |
| 3387 | types: {conditionType, resultType, resultType}, |
| 3388 | loc: parser.getNameLoc(), result&: result.operands); |
| 3389 | } |
| 3390 | |
| 3391 | void arith::SelectOp::print(OpAsmPrinter &p) { |
| 3392 | p << " " << getOperands(); |
| 3393 | p.printOptionalAttrDict(attrs: (*this)->getDiscardableAttrDictionary().getValue()); |
| 3394 | p << " : " ; |
| 3395 | if (ShapedType condType = dyn_cast<ShapedType>(Val: getCondition().getType())) |
| 3396 | p << condType << ", " ; |
| 3397 | p << getType(); |
| 3398 | } |
| 3399 | |
| 3400 | LogicalResult arith::SelectOp::verify() { |
| 3401 | Type conditionType = getCondition().getType(); |
| 3402 | if (conditionType.isSignlessInteger(width: 1)) |
| 3403 | return success(); |
| 3404 | |
| 3405 | // If the result type is a vector or tensor, the type can be a mask with the |
| 3406 | // same elements. |
| 3407 | Type resultType = getType(); |
| 3408 | if (!llvm::isa<TensorType, VectorType>(Val: resultType)) |
| 3409 | return emitOpError() << "expected condition to be a signless i1, but got " |
| 3410 | << conditionType; |
| 3411 | Type shapedConditionType = getI1SameShape(type: resultType); |
| 3412 | if (conditionType != shapedConditionType) { |
| 3413 | return emitOpError() << "expected condition type to have the same shape " |
| 3414 | "as the result type, expected " |
| 3415 | << shapedConditionType << ", but got " |
| 3416 | << conditionType; |
| 3417 | } |
| 3418 | return success(); |
| 3419 | } |
| 3420 | //===----------------------------------------------------------------------===// |
| 3421 | // ShLIOp |
| 3422 | //===----------------------------------------------------------------------===// |
| 3423 | |
| 3424 | OpFoldResult arith::ShLIOp::fold(FoldAdaptor adaptor) { |
| 3425 | // TODO: shli(x, c) -> poison when c is out of range (c >= bit width). An |
| 3426 | // out-of-range shift amount is undefined behaviour and could fold to poison, |
| 3427 | // but that would make the arith dialect depend on the ub dialect to |
| 3428 | // materialize ub.poison; left out for now. |
| 3429 | |
| 3430 | // shli(x, 0) -> x |
| 3431 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 3432 | return getLhs(); |
| 3433 | // shli(0, x) -> 0. An out-of-range shift amount yields poison, so refining |
| 3434 | // it to 0 is valid. |
| 3435 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 3436 | return getLhs(); |
| 3437 | // Don't fold if shifting more or equal than the bit width. |
| 3438 | bool bounded = false; |
| 3439 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 3440 | operands: adaptor.getOperands(), calculate: [&](const APInt &a, const APInt &b) { |
| 3441 | bounded = b.ult(RHS: b.getBitWidth()); |
| 3442 | return a.shl(ShiftAmt: b); |
| 3443 | }); |
| 3444 | return bounded ? result : Attribute(); |
| 3445 | } |
| 3446 | |
| 3447 | //===----------------------------------------------------------------------===// |
| 3448 | // ShRUIOp |
| 3449 | //===----------------------------------------------------------------------===// |
| 3450 | |
| 3451 | OpFoldResult arith::ShRUIOp::fold(FoldAdaptor adaptor) { |
| 3452 | // TODO: shrui(x, c) -> poison when c is out of range (c >= bit width). An |
| 3453 | // out-of-range shift amount is undefined behaviour and could fold to poison, |
| 3454 | // but that would make the arith dialect depend on the ub dialect to |
| 3455 | // materialize ub.poison; left out for now. |
| 3456 | |
| 3457 | // shrui(x, 0) -> x |
| 3458 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 3459 | return getLhs(); |
| 3460 | // shrui(0, x) -> 0. An out-of-range shift amount yields poison, so refining |
| 3461 | // it to 0 is valid. |
| 3462 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 3463 | return getLhs(); |
| 3464 | // shrui(x, x) -> 0. For any in-range shift amount v < bitwidth, v >> v == 0 |
| 3465 | // (v < 2^v); out-of-range amounts yield poison, so 0 is a valid refinement. |
| 3466 | if (getLhs() == getRhs()) |
| 3467 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 3468 | // Don't fold if shifting more or equal than the bit width. |
| 3469 | bool bounded = false; |
| 3470 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 3471 | operands: adaptor.getOperands(), calculate: [&](const APInt &a, const APInt &b) { |
| 3472 | bounded = b.ult(RHS: b.getBitWidth()); |
| 3473 | return a.lshr(ShiftAmt: b); |
| 3474 | }); |
| 3475 | return bounded ? result : Attribute(); |
| 3476 | } |
| 3477 | |
| 3478 | //===----------------------------------------------------------------------===// |
| 3479 | // ShRSIOp |
| 3480 | //===----------------------------------------------------------------------===// |
| 3481 | |
| 3482 | OpFoldResult arith::ShRSIOp::fold(FoldAdaptor adaptor) { |
| 3483 | // TODO: shrsi(x, c) -> poison when c is out of range (c >= bit width). An |
| 3484 | // out-of-range shift amount is undefined behaviour and could fold to poison, |
| 3485 | // but that would make the arith dialect depend on the ub dialect to |
| 3486 | // materialize ub.poison; left out for now. |
| 3487 | |
| 3488 | // shrsi(x, 0) -> x |
| 3489 | if (matchPattern(attr: adaptor.getRhs(), pattern: m_Zero())) |
| 3490 | return getLhs(); |
| 3491 | // shrsi(0, x) -> 0. An out-of-range shift amount yields poison, so refining |
| 3492 | // it to 0 is valid. |
| 3493 | if (matchPattern(attr: adaptor.getLhs(), pattern: m_Zero())) |
| 3494 | return getLhs(); |
| 3495 | // shrsi(x, x) -> 0. For any in-range shift amount v < bitwidth, v is a small |
| 3496 | // non-negative value and v >> v == 0; out-of-range amounts yield poison. |
| 3497 | if (getLhs() == getRhs()) |
| 3498 | return getIntegerAttrOfType(type: getType(), value: 0); |
| 3499 | // shrsi(-1, x) -> -1. Arithmetic shift of all-ones is all-ones for any |
| 3500 | // in-range amount; out-of-range amounts yield poison. |
| 3501 | if (APInt val; |
| 3502 | matchPattern(attr: adaptor.getLhs(), pattern: m_ConstantInt(bind_value: &val)) && val.isAllOnes()) |
| 3503 | return getLhs(); |
| 3504 | // Don't fold if shifting more or equal than the bit width. |
| 3505 | bool bounded = false; |
| 3506 | auto result = constFoldBinaryOp<IntegerAttr>( |
| 3507 | operands: adaptor.getOperands(), calculate: [&](const APInt &a, const APInt &b) { |
| 3508 | bounded = b.ult(RHS: b.getBitWidth()); |
| 3509 | return a.ashr(ShiftAmt: b); |
| 3510 | }); |
| 3511 | return bounded ? result : Attribute(); |
| 3512 | } |
| 3513 | |
| 3514 | //===----------------------------------------------------------------------===// |
| 3515 | // Atomic Enum |
| 3516 | //===----------------------------------------------------------------------===// |
| 3517 | |
| 3518 | /// Returns the identity value attribute associated with an AtomicRMWKind op. |
| 3519 | TypedAttr mlir::arith::getIdentityValueAttr(AtomicRMWKind kind, Type resultType, |
| 3520 | OpBuilder &builder, Location loc, |
| 3521 | bool useOnlyFiniteValue) { |
| 3522 | switch (kind) { |
| 3523 | case AtomicRMWKind::maximumf: { |
| 3524 | const llvm::fltSemantics &semantic = |
| 3525 | llvm::cast<FloatType>(Val&: resultType).getFloatSemantics(); |
| 3526 | APFloat identity = useOnlyFiniteValue |
| 3527 | ? APFloat::getLargest(Sem: semantic, /*Negative=*/true) |
| 3528 | : APFloat::getInf(Sem: semantic, /*Negative=*/true); |
| 3529 | return builder.getFloatAttr(type: resultType, value: identity); |
| 3530 | } |
| 3531 | case AtomicRMWKind::maxnumf: { |
| 3532 | const llvm::fltSemantics &semantic = |
| 3533 | llvm::cast<FloatType>(Val&: resultType).getFloatSemantics(); |
| 3534 | APFloat identity = useOnlyFiniteValue |
| 3535 | ? APFloat::getLargest(Sem: semantic, /*Negative=*/true) |
| 3536 | : APFloat::getNaN(Sem: semantic, /*Negative=*/true); |
| 3537 | return builder.getFloatAttr(type: resultType, value: identity); |
| 3538 | } |
| 3539 | case AtomicRMWKind::addf: |
| 3540 | case AtomicRMWKind::addi: |
| 3541 | case AtomicRMWKind::maxu: |
| 3542 | case AtomicRMWKind::ori: |
| 3543 | case AtomicRMWKind::xori: |
| 3544 | return builder.getZeroAttr(type: resultType); |
| 3545 | case AtomicRMWKind::andi: |
| 3546 | return builder.getIntegerAttr( |
| 3547 | type: resultType, |
| 3548 | value: APInt::getAllOnes(numBits: llvm::cast<IntegerType>(Val&: resultType).getWidth())); |
| 3549 | case AtomicRMWKind::maxs: |
| 3550 | return builder.getIntegerAttr( |
| 3551 | type: resultType, value: APInt::getSignedMinValue( |
| 3552 | numBits: llvm::cast<IntegerType>(Val&: resultType).getWidth())); |
| 3553 | case AtomicRMWKind::minimumf: { |
| 3554 | const llvm::fltSemantics &semantic = |
| 3555 | llvm::cast<FloatType>(Val&: resultType).getFloatSemantics(); |
| 3556 | APFloat identity = useOnlyFiniteValue |
| 3557 | ? APFloat::getLargest(Sem: semantic, /*Negative=*/false) |
| 3558 | : APFloat::getInf(Sem: semantic, /*Negative=*/false); |
| 3559 | |
| 3560 | return builder.getFloatAttr(type: resultType, value: identity); |
| 3561 | } |
| 3562 | case AtomicRMWKind::minnumf: { |
| 3563 | const llvm::fltSemantics &semantic = |
| 3564 | llvm::cast<FloatType>(Val&: resultType).getFloatSemantics(); |
| 3565 | APFloat identity = useOnlyFiniteValue |
| 3566 | ? APFloat::getLargest(Sem: semantic, /*Negative=*/false) |
| 3567 | : APFloat::getNaN(Sem: semantic, /*Negative=*/false); |
| 3568 | return builder.getFloatAttr(type: resultType, value: identity); |
| 3569 | } |
| 3570 | case AtomicRMWKind::mins: |
| 3571 | return builder.getIntegerAttr( |
| 3572 | type: resultType, value: APInt::getSignedMaxValue( |
| 3573 | numBits: llvm::cast<IntegerType>(Val&: resultType).getWidth())); |
| 3574 | case AtomicRMWKind::minu: |
| 3575 | return builder.getIntegerAttr( |
| 3576 | type: resultType, |
| 3577 | value: APInt::getMaxValue(numBits: llvm::cast<IntegerType>(Val&: resultType).getWidth())); |
| 3578 | case AtomicRMWKind::muli: |
| 3579 | return builder.getIntegerAttr(type: resultType, value: 1); |
| 3580 | case AtomicRMWKind::mulf: |
| 3581 | return builder.getFloatAttr(type: resultType, value: 1); |
| 3582 | // `assign` is not a reduction and has no identity element. |
| 3583 | case AtomicRMWKind::assign: |
| 3584 | break; |
| 3585 | } |
| 3586 | (void)emitOptionalError(loc, args: "Reduction operation type not supported" ); |
| 3587 | return nullptr; |
| 3588 | } |
| 3589 | |
| 3590 | /// Returns the identity numeric value of the given op. |
| 3591 | std::optional<TypedAttr> mlir::arith::getNeutralElement(Operation *op) { |
| 3592 | std::optional<AtomicRMWKind> maybeKind = |
| 3593 | llvm::TypeSwitch<Operation *, std::optional<AtomicRMWKind>>(op) |
| 3594 | // Floating-point operations. |
| 3595 | .Case(caseFn: [](arith::AddFOp op) { return AtomicRMWKind::addf; }) |
| 3596 | .Case(caseFn: [](arith::MulFOp op) { return AtomicRMWKind::mulf; }) |
| 3597 | .Case(caseFn: [](arith::MaximumFOp op) { return AtomicRMWKind::maximumf; }) |
| 3598 | .Case(caseFn: [](arith::MinimumFOp op) { return AtomicRMWKind::minimumf; }) |
| 3599 | .Case(caseFn: [](arith::MaxNumFOp op) { return AtomicRMWKind::maxnumf; }) |
| 3600 | .Case(caseFn: [](arith::MinNumFOp op) { return AtomicRMWKind::minnumf; }) |
| 3601 | // Integer operations. |
| 3602 | .Case(caseFn: [](arith::AddIOp op) { return AtomicRMWKind::addi; }) |
| 3603 | .Case(caseFn: [](arith::OrIOp op) { return AtomicRMWKind::ori; }) |
| 3604 | .Case(caseFn: [](arith::XOrIOp op) { return AtomicRMWKind::xori; }) |
| 3605 | .Case(caseFn: [](arith::AndIOp op) { return AtomicRMWKind::andi; }) |
| 3606 | .Case(caseFn: [](arith::MaxUIOp op) { return AtomicRMWKind::maxu; }) |
| 3607 | .Case(caseFn: [](arith::MinUIOp op) { return AtomicRMWKind::minu; }) |
| 3608 | .Case(caseFn: [](arith::MaxSIOp op) { return AtomicRMWKind::maxs; }) |
| 3609 | .Case(caseFn: [](arith::MinSIOp op) { return AtomicRMWKind::mins; }) |
| 3610 | .Case(caseFn: [](arith::MulIOp op) { return AtomicRMWKind::muli; }) |
| 3611 | .Default(std::nullopt); |
| 3612 | if (!maybeKind) { |
| 3613 | return std::nullopt; |
| 3614 | } |
| 3615 | |
| 3616 | bool useOnlyFiniteValue = false; |
| 3617 | auto fmfOpInterface = dyn_cast<ArithFastMathInterface>(Val: op); |
| 3618 | if (fmfOpInterface) { |
| 3619 | arith::FastMathFlagsAttr fmfAttr = fmfOpInterface.getFastMathFlagsAttr(); |
| 3620 | useOnlyFiniteValue = |
| 3621 | bitEnumContainsAny(bits: fmfAttr.getValue(), bit: arith::FastMathFlags::ninf); |
| 3622 | if (*maybeKind == AtomicRMWKind::maxnumf || |
| 3623 | *maybeKind == AtomicRMWKind::minnumf) { |
| 3624 | if (bitEnumContainsAny(bits: fmfAttr.getValue(), bit: arith::FastMathFlags::nnan)) { |
| 3625 | *maybeKind = *maybeKind == AtomicRMWKind::maxnumf |
| 3626 | ? AtomicRMWKind::maximumf |
| 3627 | : AtomicRMWKind::minimumf; |
| 3628 | } else { |
| 3629 | useOnlyFiniteValue = false; |
| 3630 | } |
| 3631 | } |
| 3632 | } |
| 3633 | |
| 3634 | // Builder only used as helper for attribute creation. |
| 3635 | OpBuilder b(op->getContext()); |
| 3636 | Type resultType = op->getResult(idx: 0).getType(); |
| 3637 | |
| 3638 | return getIdentityValueAttr(kind: *maybeKind, resultType, builder&: b, loc: op->getLoc(), |
| 3639 | useOnlyFiniteValue); |
| 3640 | } |
| 3641 | |
| 3642 | /// Returns the identity value associated with an AtomicRMWKind op. |
| 3643 | Value mlir::arith::getIdentityValue(AtomicRMWKind op, Type resultType, |
| 3644 | OpBuilder &builder, Location loc, |
| 3645 | bool useOnlyFiniteValue) { |
| 3646 | if (auto attr = getIdentityValueAttr(kind: op, resultType, builder, loc, |
| 3647 | useOnlyFiniteValue)) |
| 3648 | return arith::ConstantOp::create(builder, location: loc, value: attr); |
| 3649 | return {}; |
| 3650 | } |
| 3651 | |
| 3652 | /// Return the value obtained by applying the reduction operation kind |
| 3653 | /// associated with a binary AtomicRMWKind op to `lhs` and `rhs`. |
| 3654 | Value mlir::arith::getReductionOp(AtomicRMWKind op, OpBuilder &builder, |
| 3655 | Location loc, Value lhs, Value rhs) { |
| 3656 | switch (op) { |
| 3657 | case AtomicRMWKind::addf: |
| 3658 | return arith::AddFOp::create(builder, location: loc, lhs, rhs); |
| 3659 | case AtomicRMWKind::addi: |
| 3660 | return arith::AddIOp::create(builder, location: loc, lhs, rhs); |
| 3661 | case AtomicRMWKind::mulf: |
| 3662 | return arith::MulFOp::create(builder, location: loc, lhs, rhs); |
| 3663 | case AtomicRMWKind::muli: |
| 3664 | return arith::MulIOp::create(builder, location: loc, lhs, rhs); |
| 3665 | case AtomicRMWKind::maximumf: |
| 3666 | return arith::MaximumFOp::create(builder, location: loc, lhs, rhs); |
| 3667 | case AtomicRMWKind::minimumf: |
| 3668 | return arith::MinimumFOp::create(builder, location: loc, lhs, rhs); |
| 3669 | case AtomicRMWKind::maxnumf: |
| 3670 | return arith::MaxNumFOp::create(builder, location: loc, lhs, rhs); |
| 3671 | case AtomicRMWKind::minnumf: |
| 3672 | return arith::MinNumFOp::create(builder, location: loc, lhs, rhs); |
| 3673 | case AtomicRMWKind::maxs: |
| 3674 | return arith::MaxSIOp::create(builder, location: loc, lhs, rhs); |
| 3675 | case AtomicRMWKind::mins: |
| 3676 | return arith::MinSIOp::create(builder, location: loc, lhs, rhs); |
| 3677 | case AtomicRMWKind::maxu: |
| 3678 | return arith::MaxUIOp::create(builder, location: loc, lhs, rhs); |
| 3679 | case AtomicRMWKind::minu: |
| 3680 | return arith::MinUIOp::create(builder, location: loc, lhs, rhs); |
| 3681 | case AtomicRMWKind::ori: |
| 3682 | return arith::OrIOp::create(builder, location: loc, lhs, rhs); |
| 3683 | case AtomicRMWKind::andi: |
| 3684 | return arith::AndIOp::create(builder, location: loc, lhs, rhs); |
| 3685 | case AtomicRMWKind::xori: |
| 3686 | return arith::XOrIOp::create(builder, location: loc, lhs, rhs); |
| 3687 | // `assign` is not a reduction and has no corresponding binary operation. |
| 3688 | case AtomicRMWKind::assign: |
| 3689 | break; |
| 3690 | } |
| 3691 | (void)emitOptionalError(loc, args: "Reduction operation type not supported" ); |
| 3692 | return nullptr; |
| 3693 | } |
| 3694 | |
| 3695 | //===----------------------------------------------------------------------===// |
| 3696 | // TableGen'd op method definitions |
| 3697 | //===----------------------------------------------------------------------===// |
| 3698 | |
| 3699 | #define GET_OP_CLASSES |
| 3700 | #include "mlir/Dialect/Arith/IR/ArithOps.cpp.inc" |
| 3701 | |
| 3702 | //===----------------------------------------------------------------------===// |
| 3703 | // TableGen'd enum attribute definitions |
| 3704 | //===----------------------------------------------------------------------===// |
| 3705 | |
| 3706 | #include "mlir/Dialect/Arith/IR/ArithOpsEnums.cpp.inc" |
| 3707 | |