| 1 | //===- IntRangeOptimizations.cpp - Optimizations based on integer ranges --===// |
| 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 <utility> |
| 10 | |
| 11 | #include "llvm/ADT/TypeSwitch.h" |
| 12 | |
| 13 | #include "mlir/Analysis/DataFlow/ConstantPropagationAnalysis.h" |
| 14 | #include "mlir/Analysis/DataFlow/Utils.h" |
| 15 | #include "mlir/Analysis/DataFlowFramework.h" |
| 16 | #include "mlir/Dialect/Arith/Transforms/Passes.h" |
| 17 | |
| 18 | #include "mlir/Analysis/DataFlow/DeadCodeAnalysis.h" |
| 19 | #include "mlir/Analysis/DataFlow/IntegerRangeAnalysis.h" |
| 20 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 21 | #include "mlir/IR/IRMapping.h" |
| 22 | #include "mlir/IR/Matchers.h" |
| 23 | #include "mlir/IR/PatternMatch.h" |
| 24 | #include "mlir/IR/TypeUtilities.h" |
| 25 | #include "mlir/Interfaces/LoopLikeInterface.h" |
| 26 | #include "mlir/Interfaces/SideEffectInterfaces.h" |
| 27 | #include "mlir/Transforms/FoldUtils.h" |
| 28 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 29 | |
| 30 | namespace mlir::arith { |
| 31 | #define GEN_PASS_DEF_ARITHINTRANGEOPTS |
| 32 | #include "mlir/Dialect/Arith/Transforms/Passes.h.inc" |
| 33 | |
| 34 | #define GEN_PASS_DEF_ARITHINTRANGENARROWING |
| 35 | #include "mlir/Dialect/Arith/Transforms/Passes.h.inc" |
| 36 | } // namespace mlir::arith |
| 37 | |
| 38 | using namespace mlir; |
| 39 | using namespace mlir::arith; |
| 40 | using namespace mlir::dataflow; |
| 41 | |
| 42 | static std::optional<APInt> getMaybeConstantValue(DataFlowSolver &solver, |
| 43 | Value value) { |
| 44 | auto *maybeInferredRange = |
| 45 | solver.lookupState<IntegerValueRangeLattice>(anchor: value); |
| 46 | if (!maybeInferredRange || maybeInferredRange->getValue().isUninitialized()) |
| 47 | return std::nullopt; |
| 48 | const ConstantIntRanges &inferredRange = |
| 49 | maybeInferredRange->getValue().getValue(); |
| 50 | return inferredRange.getConstantValue(); |
| 51 | } |
| 52 | |
| 53 | static void copyIntegerRange(DataFlowSolver &solver, Value oldVal, |
| 54 | Value newVal) { |
| 55 | auto *oldState = solver.lookupState<IntegerValueRangeLattice>(anchor: oldVal); |
| 56 | if (!oldState) |
| 57 | return; |
| 58 | (void)solver.getOrCreateState<IntegerValueRangeLattice>(anchor: newVal)->join( |
| 59 | rhs: *oldState); |
| 60 | } |
| 61 | |
| 62 | namespace mlir::dataflow { |
| 63 | /// Patterned after SCCP |
| 64 | LogicalResult maybeReplaceWithConstant(DataFlowSolver &solver, |
| 65 | RewriterBase &rewriter, Value value) { |
| 66 | if (value.use_empty()) |
| 67 | return failure(); |
| 68 | std::optional<APInt> maybeConstValue = getMaybeConstantValue(solver, value); |
| 69 | if (!maybeConstValue.has_value()) |
| 70 | return failure(); |
| 71 | |
| 72 | Type type = value.getType(); |
| 73 | // If the type or element type is non-integral, the attribute constructor |
| 74 | // will crash, so eagerly check for an integer type to avoid this. |
| 75 | if (!getElementTypeOrSelf(type).isIntOrIndex()) |
| 76 | return failure(); |
| 77 | |
| 78 | // Bail out if the inferred APInt bitwidth does not match the storage width |
| 79 | // of the IR type; IntegerAttr::get would assert otherwise. |
| 80 | unsigned storageWidth = ConstantIntRanges::getStorageBitwidth(type); |
| 81 | if (storageWidth != 0 && maybeConstValue->getBitWidth() != storageWidth) |
| 82 | return failure(); |
| 83 | |
| 84 | Location loc = value.getLoc(); |
| 85 | Operation *maybeDefiningOp = value.getDefiningOp(); |
| 86 | Dialect *valueDialect = |
| 87 | maybeDefiningOp ? maybeDefiningOp->getDialect() |
| 88 | : value.getParentRegion()->getParentOp()->getDialect(); |
| 89 | |
| 90 | Attribute constAttr; |
| 91 | if (auto shaped = dyn_cast<ShapedType>(Val&: type)) { |
| 92 | constAttr = mlir::DenseIntElementsAttr::get(type: shaped, arg&: *maybeConstValue); |
| 93 | } else { |
| 94 | constAttr = rewriter.getIntegerAttr(type, value: *maybeConstValue); |
| 95 | } |
| 96 | Operation *constOp = |
| 97 | valueDialect->materializeConstant(builder&: rewriter, value: constAttr, type, loc); |
| 98 | // Fall back to arith.constant if the dialect materializer doesn't know what |
| 99 | // to do with an integer constant. |
| 100 | if (!constOp) |
| 101 | constOp = rewriter.getContext() |
| 102 | ->getLoadedDialect<ArithDialect>() |
| 103 | ->materializeConstant(builder&: rewriter, value: constAttr, type, loc); |
| 104 | if (!constOp) |
| 105 | return failure(); |
| 106 | |
| 107 | OpResult res = constOp->getResult(idx: 0); |
| 108 | if (solver.lookupState<dataflow::IntegerValueRangeLattice>(anchor: res)) |
| 109 | solver.eraseState(anchor: res); |
| 110 | copyIntegerRange(solver, oldVal: value, newVal: res); |
| 111 | rewriter.replaceAllUsesWith(from: value, to: res); |
| 112 | return success(); |
| 113 | } |
| 114 | } // namespace mlir::dataflow |
| 115 | |
| 116 | namespace { |
| 117 | class DataFlowListener : public RewriterBase::Listener { |
| 118 | public: |
| 119 | DataFlowListener(DataFlowSolver &s) : s(s) {} |
| 120 | |
| 121 | protected: |
| 122 | void notifyOperationErased(Operation *op) override { |
| 123 | s.eraseState(anchor: s.getProgramPointAfter(op)); |
| 124 | for (Value res : op->getResults()) |
| 125 | s.eraseState(anchor: res); |
| 126 | } |
| 127 | |
| 128 | DataFlowSolver &s; |
| 129 | }; |
| 130 | |
| 131 | /// Rewrite any results of `op` that were inferred to be constant integers to |
| 132 | /// and replace their uses with that constant. Return success() if all results |
| 133 | /// where thus replaced and the operation is erased. Also replace any block |
| 134 | /// arguments with their constant values. |
| 135 | struct MaterializeKnownConstantValues : public RewritePattern { |
| 136 | MaterializeKnownConstantValues(MLIRContext *context, DataFlowSolver &s) |
| 137 | : RewritePattern::RewritePattern(Pattern::MatchAnyOpTypeTag(), |
| 138 | /*benefit=*/1, context), |
| 139 | solver(s) {} |
| 140 | |
| 141 | LogicalResult matchAndRewrite(Operation *op, |
| 142 | PatternRewriter &rewriter) const override { |
| 143 | if (matchPattern(op, pattern: m_Constant())) |
| 144 | return failure(); |
| 145 | |
| 146 | // We need to check isIntOrIndex() and APInt bitwidth compatibility here |
| 147 | // as well to avoid infinite loops in the greedy pattern rewriter. If we |
| 148 | // only check in maybeReplaceWithConstant, this lambda might still return |
| 149 | // true for values that cannot be materialized, causing the pattern to |
| 150 | // match and claim success without making any changes, leading to |
| 151 | // non-convergence. |
| 152 | auto needsReplacing = [&](Value v) { |
| 153 | if (!getElementTypeOrSelf(type: v.getType()).isIntOrIndex()) |
| 154 | return false; |
| 155 | std::optional<APInt> maybeConstValue = getMaybeConstantValue(solver, value: v); |
| 156 | if (!maybeConstValue.has_value() || v.use_empty()) |
| 157 | return false; |
| 158 | unsigned storageWidth = |
| 159 | ConstantIntRanges::getStorageBitwidth(type: v.getType()); |
| 160 | return storageWidth == 0 || |
| 161 | maybeConstValue->getBitWidth() == storageWidth; |
| 162 | }; |
| 163 | bool hasConstantResults = llvm::any_of(Range: op->getResults(), P: needsReplacing); |
| 164 | if (op->getNumRegions() == 0) |
| 165 | if (!hasConstantResults) |
| 166 | return failure(); |
| 167 | bool hasConstantRegionArgs = false; |
| 168 | for (Region ®ion : op->getRegions()) { |
| 169 | for (Block &block : region.getBlocks()) { |
| 170 | hasConstantRegionArgs |= |
| 171 | llvm::any_of(Range: block.getArguments(), P: needsReplacing); |
| 172 | } |
| 173 | } |
| 174 | if (!hasConstantResults && !hasConstantRegionArgs) |
| 175 | return failure(); |
| 176 | |
| 177 | bool replacedAll = (op->getNumResults() != 0); |
| 178 | for (Value v : op->getResults()) |
| 179 | replacedAll &= |
| 180 | (succeeded(Result: maybeReplaceWithConstant(solver, rewriter, value: v)) || |
| 181 | v.use_empty()); |
| 182 | if (replacedAll && isOpTriviallyDead(op)) { |
| 183 | rewriter.eraseOp(op); |
| 184 | return success(); |
| 185 | } |
| 186 | |
| 187 | PatternRewriter::InsertionGuard guard(rewriter); |
| 188 | for (Region ®ion : op->getRegions()) { |
| 189 | for (Block &block : region.getBlocks()) { |
| 190 | rewriter.setInsertionPointToStart(&block); |
| 191 | for (BlockArgument &arg : block.getArguments()) { |
| 192 | (void)maybeReplaceWithConstant(solver, rewriter, value: arg); |
| 193 | } |
| 194 | } |
| 195 | } |
| 196 | |
| 197 | return success(); |
| 198 | } |
| 199 | |
| 200 | private: |
| 201 | DataFlowSolver &solver; |
| 202 | }; |
| 203 | |
| 204 | template <typename RemOp> |
| 205 | struct DeleteTrivialRem : public OpRewritePattern<RemOp> { |
| 206 | DeleteTrivialRem(MLIRContext *context, DataFlowSolver &s) |
| 207 | : OpRewritePattern<RemOp>(context), solver(s) {} |
| 208 | |
| 209 | LogicalResult matchAndRewrite(RemOp op, |
| 210 | PatternRewriter &rewriter) const override { |
| 211 | Value lhs = op.getOperand(0); |
| 212 | Value rhs = op.getOperand(1); |
| 213 | // TODO: Support index types once integer range inference can use the target |
| 214 | // index bitwidth, e.g. from DLTI. |
| 215 | if (isa<IndexType>(Val: getElementTypeOrSelf(type: lhs.getType()))) |
| 216 | return failure(); |
| 217 | APInt modulus; |
| 218 | bool isUnsigned = isa<RemUIOp>(op); |
| 219 | // Any nonzero bit pattern is a valid unsigned modulus. Keep the existing |
| 220 | // positive-modulus restriction for signed remainder. |
| 221 | if (!matchPattern(value: rhs, pattern: m_ConstantInt(bind_value: &modulus)) || modulus.isZero() || |
| 222 | (!isUnsigned && modulus.isNegative())) |
| 223 | return failure(); |
| 224 | auto *maybeLhsRange = solver.lookupState<IntegerValueRangeLattice>(anchor: lhs); |
| 225 | if (!maybeLhsRange || maybeLhsRange->getValue().isUninitialized()) |
| 226 | return failure(); |
| 227 | const ConstantIntRanges &lhsRange = maybeLhsRange->getValue().getValue(); |
| 228 | const APInt &min = isUnsigned ? lhsRange.umin() : lhsRange.smin(); |
| 229 | const APInt &max = isUnsigned ? lhsRange.umax() : lhsRange.smax(); |
| 230 | if (min.getBitWidth() != modulus.getBitWidth() || |
| 231 | max.getBitWidth() != modulus.getBitWidth()) |
| 232 | return failure(); |
| 233 | // The minima and maxima here are given as closed ranges, we must be |
| 234 | // non-negative for signed remainder and strictly less than the modulus. |
| 235 | if ((!isUnsigned && min.isNegative()) || min.uge(RHS: modulus)) |
| 236 | return failure(); |
| 237 | if ((!isUnsigned && max.isNegative()) || max.uge(RHS: modulus)) |
| 238 | return failure(); |
| 239 | if (!min.ule(RHS: max)) |
| 240 | return failure(); |
| 241 | |
| 242 | // With all those conditions out of the way, we know thas this invocation of |
| 243 | // a remainder is a noop because the input is strictly within the range |
| 244 | // [0, modulus), so get rid of it. |
| 245 | rewriter.replaceOp(op, ValueRange{lhs}); |
| 246 | return success(); |
| 247 | } |
| 248 | |
| 249 | private: |
| 250 | DataFlowSolver &solver; |
| 251 | }; |
| 252 | |
| 253 | /// Gather ranges for all the values in `values`. Appends to the existing |
| 254 | /// vector. |
| 255 | static LogicalResult collectRanges(DataFlowSolver &solver, ValueRange values, |
| 256 | SmallVectorImpl<ConstantIntRanges> &ranges) { |
| 257 | for (Value val : values) { |
| 258 | auto *maybeInferredRange = |
| 259 | solver.lookupState<IntegerValueRangeLattice>(anchor: val); |
| 260 | if (!maybeInferredRange || maybeInferredRange->getValue().isUninitialized()) |
| 261 | return failure(); |
| 262 | |
| 263 | const ConstantIntRanges &inferredRange = |
| 264 | maybeInferredRange->getValue().getValue(); |
| 265 | ranges.push_back(Elt: inferredRange); |
| 266 | } |
| 267 | return success(); |
| 268 | } |
| 269 | |
| 270 | /// Return int type truncated to `targetBitwidth`. If `srcType` is shaped, |
| 271 | /// return shaped type as well. |
| 272 | static Type getTargetType(Type srcType, unsigned targetBitwidth) { |
| 273 | auto dstType = IntegerType::get(context: srcType.getContext(), width: targetBitwidth); |
| 274 | if (auto shaped = dyn_cast<ShapedType>(Val&: srcType)) |
| 275 | return shaped.clone(elementType: dstType); |
| 276 | |
| 277 | assert(srcType.isIntOrIndex() && "Invalid src type" ); |
| 278 | return dstType; |
| 279 | } |
| 280 | |
| 281 | namespace { |
| 282 | // Enum for tracking which type of truncation should be performed |
| 283 | // to narrow an operation, if any. |
| 284 | enum class CastKind : uint8_t { None, Signed, Unsigned, Both }; |
| 285 | } // namespace |
| 286 | |
| 287 | /// If the values within `range` can be represented using only `width` bits, |
| 288 | /// return the kind of truncation needed to preserve that property. |
| 289 | /// |
| 290 | /// This check relies on the fact that the signed and unsigned ranges are both |
| 291 | /// always correct, but that one might be an approximation of the other, |
| 292 | /// so we want to use the correct truncation operation. |
| 293 | static CastKind checkTruncatability(const ConstantIntRanges &range, |
| 294 | unsigned targetWidth) { |
| 295 | unsigned srcWidth = range.smin().getBitWidth(); |
| 296 | if (srcWidth <= targetWidth) |
| 297 | return CastKind::None; |
| 298 | unsigned removedWidth = srcWidth - targetWidth; |
| 299 | // The sign bits need to extend into the sign bit of the target width. For |
| 300 | // example, if we're truncating 64 bits to 32, we need 64 - 32 + 1 = 33 sign |
| 301 | // bits. |
| 302 | bool canTruncateSigned = |
| 303 | range.smin().getNumSignBits() >= (removedWidth + 1) && |
| 304 | range.smax().getNumSignBits() >= (removedWidth + 1); |
| 305 | bool canTruncateUnsigned = range.umin().countLeadingZeros() >= removedWidth && |
| 306 | range.umax().countLeadingZeros() >= removedWidth; |
| 307 | if (canTruncateSigned && canTruncateUnsigned) |
| 308 | return CastKind::Both; |
| 309 | if (canTruncateSigned) |
| 310 | return CastKind::Signed; |
| 311 | if (canTruncateUnsigned) |
| 312 | return CastKind::Unsigned; |
| 313 | return CastKind::None; |
| 314 | } |
| 315 | |
| 316 | static CastKind mergeCastKinds(CastKind lhs, CastKind rhs) { |
| 317 | if (lhs == CastKind::None || rhs == CastKind::None) |
| 318 | return CastKind::None; |
| 319 | if (lhs == CastKind::Both) |
| 320 | return rhs; |
| 321 | if (rhs == CastKind::Both) |
| 322 | return lhs; |
| 323 | if (lhs == rhs) |
| 324 | return lhs; |
| 325 | return CastKind::None; |
| 326 | } |
| 327 | |
| 328 | static Value doCast(OpBuilder &builder, Location loc, Value src, Type dstType, |
| 329 | CastKind castKind) { |
| 330 | Type srcType = src.getType(); |
| 331 | assert(isa<VectorType>(srcType) == isa<VectorType>(dstType) && |
| 332 | "Mixing vector and non-vector types" ); |
| 333 | assert(castKind != CastKind::None && "Can't cast when casting isn't allowed" ); |
| 334 | Type srcElemType = getElementTypeOrSelf(type: srcType); |
| 335 | Type dstElemType = getElementTypeOrSelf(type: dstType); |
| 336 | assert(srcElemType.isIntOrIndex() && "Invalid src type" ); |
| 337 | assert(dstElemType.isIntOrIndex() && "Invalid dst type" ); |
| 338 | if (srcType == dstType) |
| 339 | return src; |
| 340 | |
| 341 | if (isa<IndexType>(Val: srcElemType) || isa<IndexType>(Val: dstElemType)) { |
| 342 | if (castKind == CastKind::Signed) |
| 343 | return arith::IndexCastOp::create(builder, location: loc, out: dstType, in: src); |
| 344 | return arith::IndexCastUIOp::create(builder, location: loc, out: dstType, in: src); |
| 345 | } |
| 346 | |
| 347 | auto srcInt = cast<IntegerType>(Val&: srcElemType); |
| 348 | auto dstInt = cast<IntegerType>(Val&: dstElemType); |
| 349 | if (dstInt.getWidth() < srcInt.getWidth()) |
| 350 | return arith::TruncIOp::create(builder, location: loc, out: dstType, in: src); |
| 351 | |
| 352 | if (castKind == CastKind::Signed) |
| 353 | return arith::ExtSIOp::create(builder, location: loc, out: dstType, in: src); |
| 354 | return arith::ExtUIOp::create(builder, location: loc, out: dstType, in: src); |
| 355 | } |
| 356 | |
| 357 | struct NarrowElementwise final : OpTraitRewritePattern<OpTrait::Elementwise> { |
| 358 | NarrowElementwise(MLIRContext *context, DataFlowSolver &s, |
| 359 | ArrayRef<unsigned> target) |
| 360 | : OpTraitRewritePattern(context), solver(s), targetBitwidths(target) {} |
| 361 | |
| 362 | using Base::Base; |
| 363 | LogicalResult matchAndRewrite(Operation *op, |
| 364 | PatternRewriter &rewriter) const override { |
| 365 | if (op->getNumResults() == 0) |
| 366 | return rewriter.notifyMatchFailure(arg&: op, msg: "can't narrow resultless op" ); |
| 367 | |
| 368 | // Inline size chosen empirically based on compilation profiling. |
| 369 | // Profiled: 2.6M calls, avg=1.7+-1.3. N=4 covers >95% of cases inline. |
| 370 | SmallVector<ConstantIntRanges, 4> ranges; |
| 371 | if (failed(Result: collectRanges(solver, values: op->getOperands(), ranges))) |
| 372 | return rewriter.notifyMatchFailure(arg&: op, msg: "input without specified range" ); |
| 373 | if (failed(Result: collectRanges(solver, values: op->getResults(), ranges))) |
| 374 | return rewriter.notifyMatchFailure(arg&: op, msg: "output without specified range" ); |
| 375 | |
| 376 | Type srcType = op->getResult(idx: 0).getType(); |
| 377 | if (!llvm::all_equal(Range: op->getResultTypes())) |
| 378 | return rewriter.notifyMatchFailure(arg&: op, msg: "mismatched result types" ); |
| 379 | if (op->getNumOperands() == 0 || |
| 380 | !llvm::all_of(Range: op->getOperandTypes(), |
| 381 | P: [=](Type t) { return t == srcType; })) |
| 382 | return rewriter.notifyMatchFailure( |
| 383 | arg&: op, msg: "no operands or operand types don't match result type" ); |
| 384 | |
| 385 | for (unsigned targetBitwidth : targetBitwidths) { |
| 386 | CastKind castKind = CastKind::Both; |
| 387 | for (const ConstantIntRanges &range : ranges) { |
| 388 | castKind = mergeCastKinds(lhs: castKind, |
| 389 | rhs: checkTruncatability(range, targetWidth: targetBitwidth)); |
| 390 | if (castKind == CastKind::None) |
| 391 | break; |
| 392 | } |
| 393 | // Constrain the cast kind to preserve the interpretation required by the |
| 394 | // operation. |
| 395 | auto castKindForOp = |
| 396 | llvm::TypeSwitch<Operation *, CastKind>(op) |
| 397 | .Case<arith::DivSIOp, arith::CeilDivSIOp, arith::FloorDivSIOp, |
| 398 | arith::RemSIOp, arith::MaxSIOp, arith::MinSIOp, |
| 399 | arith::ShRSIOp>(caseFn: [](auto) { return CastKind::Signed; }) |
| 400 | .Case<arith::DivUIOp, arith::CeilDivUIOp, arith::RemUIOp, |
| 401 | arith::ShRUIOp>(caseFn: [](auto) { return CastKind::Unsigned; }) |
| 402 | .Default(defaultResult: CastKind::Both); |
| 403 | castKind = mergeCastKinds(lhs: castKind, rhs: castKindForOp); |
| 404 | if (castKind == CastKind::None) |
| 405 | continue; |
| 406 | // A shift by an amount >= the bitwidth is poison, so only narrow shifts |
| 407 | // when the shift amount (second operand) stays below the target width. |
| 408 | if (isa<arith::ShLIOp, arith::ShRSIOp, arith::ShRUIOp>(Val: op) && |
| 409 | !ranges[1].umax().ult(RHS: targetBitwidth)) |
| 410 | continue; |
| 411 | Type targetType = getTargetType(srcType, targetBitwidth); |
| 412 | if (targetType == srcType) |
| 413 | continue; |
| 414 | |
| 415 | Location loc = op->getLoc(); |
| 416 | IRMapping mapping; |
| 417 | for (auto [arg, argRange] : llvm::zip_first(t: op->getOperands(), u&: ranges)) { |
| 418 | CastKind argCastKind = castKind; |
| 419 | // When dealing with `index` values, preserve non-negativity in the |
| 420 | // index_casts since we can't recover this in unsigned when equivalent. |
| 421 | if (argCastKind == CastKind::Signed && argRange.smin().isNonNegative()) |
| 422 | argCastKind = CastKind::Both; |
| 423 | Value newArg = doCast(builder&: rewriter, loc, src: arg, dstType: targetType, castKind: argCastKind); |
| 424 | mapping.map(from: arg, to: newArg); |
| 425 | } |
| 426 | |
| 427 | Operation *newOp = rewriter.clone(op&: *op, mapper&: mapping); |
| 428 | rewriter.modifyOpInPlace(root: newOp, callable: [&]() { |
| 429 | for (OpResult res : newOp->getResults()) { |
| 430 | res.setType(targetType); |
| 431 | } |
| 432 | // Retain an overflow flag only when the casts preserve operand and |
| 433 | // result values in the flag's signed or unsigned interpretation. |
| 434 | if (auto overflowOp = |
| 435 | dyn_cast<arith::ArithIntegerOverflowFlagsInterface>(Val: newOp)) { |
| 436 | auto flags = overflowOp.getOverflowAttr().getValue(); |
| 437 | if (castKind == CastKind::Unsigned) |
| 438 | flags = bitEnumClear(bits: flags, bit: IntegerOverflowFlags::nsw); |
| 439 | else if (castKind == CastKind::Signed) |
| 440 | flags = bitEnumClear(bits: flags, bit: IntegerOverflowFlags::nuw); |
| 441 | newOp->setAttr( |
| 442 | name: overflowOp.getIntegerOverflowAttrName(), |
| 443 | value: IntegerOverflowFlagsAttr::get(context: rewriter.getContext(), value: flags)); |
| 444 | } |
| 445 | }); |
| 446 | SmallVector<Value> newResults; |
| 447 | for (auto [newRes, oldRes] : |
| 448 | llvm::zip_equal(t: newOp->getResults(), u: op->getResults())) { |
| 449 | Value castBack = doCast(builder&: rewriter, loc, src: newRes, dstType: srcType, castKind); |
| 450 | copyIntegerRange(solver, oldVal: oldRes, newVal: castBack); |
| 451 | newResults.push_back(Elt: castBack); |
| 452 | } |
| 453 | |
| 454 | rewriter.replaceOp(op, newValues: newResults); |
| 455 | return success(); |
| 456 | } |
| 457 | return failure(); |
| 458 | } |
| 459 | |
| 460 | private: |
| 461 | DataFlowSolver &solver; |
| 462 | SmallVector<unsigned, 4> targetBitwidths; |
| 463 | }; |
| 464 | |
| 465 | struct NarrowCmpI final : OpRewritePattern<arith::CmpIOp> { |
| 466 | NarrowCmpI(MLIRContext *context, DataFlowSolver &s, ArrayRef<unsigned> target) |
| 467 | : OpRewritePattern(context), solver(s), targetBitwidths(target) {} |
| 468 | |
| 469 | LogicalResult matchAndRewrite(arith::CmpIOp op, |
| 470 | PatternRewriter &rewriter) const override { |
| 471 | Value lhs = op.getLhs(); |
| 472 | Value rhs = op.getRhs(); |
| 473 | |
| 474 | SmallVector<ConstantIntRanges> ranges; |
| 475 | if (failed(Result: collectRanges(solver, values: op.getOperands(), ranges))) |
| 476 | return failure(); |
| 477 | const ConstantIntRanges &lhsRange = ranges[0]; |
| 478 | const ConstantIntRanges &rhsRange = ranges[1]; |
| 479 | |
| 480 | auto isSignedCmpPredicate = [](arith::CmpIPredicate pred) -> bool { |
| 481 | return pred == arith::CmpIPredicate::sge || |
| 482 | pred == arith::CmpIPredicate::sgt || |
| 483 | pred == arith::CmpIPredicate::sle || |
| 484 | pred == arith::CmpIPredicate::slt; |
| 485 | }; |
| 486 | // If we're to narrow the input values via a cast, we should preserve the |
| 487 | // sign. |
| 488 | CastKind predicateBasedCastRestriction = |
| 489 | isSignedCmpPredicate(op.getPredicate()) ? CastKind::Signed |
| 490 | : CastKind::Both; |
| 491 | |
| 492 | Type srcType = lhs.getType(); |
| 493 | for (unsigned targetBitwidth : targetBitwidths) { |
| 494 | CastKind lhsCastKind = checkTruncatability(range: lhsRange, targetWidth: targetBitwidth); |
| 495 | CastKind rhsCastKind = checkTruncatability(range: rhsRange, targetWidth: targetBitwidth); |
| 496 | CastKind castKind = mergeCastKinds(lhs: lhsCastKind, rhs: rhsCastKind); |
| 497 | castKind = mergeCastKinds(lhs: castKind, rhs: predicateBasedCastRestriction); |
| 498 | // Note: this includes target width > src width, as well as the unsigned |
| 499 | // truncatability & signed predicate scenario. |
| 500 | if (castKind == CastKind::None) |
| 501 | continue; |
| 502 | |
| 503 | Type targetType = getTargetType(srcType, targetBitwidth); |
| 504 | if (targetType == srcType) |
| 505 | continue; |
| 506 | |
| 507 | Location loc = op->getLoc(); |
| 508 | IRMapping mapping; |
| 509 | Value lhsCast = doCast(builder&: rewriter, loc, src: lhs, dstType: targetType, castKind: lhsCastKind); |
| 510 | Value rhsCast = doCast(builder&: rewriter, loc, src: rhs, dstType: targetType, castKind: rhsCastKind); |
| 511 | mapping.map(from: lhs, to: lhsCast); |
| 512 | mapping.map(from: rhs, to: rhsCast); |
| 513 | |
| 514 | Operation *newOp = rewriter.clone(op&: *op, mapper&: mapping); |
| 515 | copyIntegerRange(solver, oldVal: op.getResult(), newVal: newOp->getResult(idx: 0)); |
| 516 | rewriter.replaceOp(op, newValues: newOp->getResults()); |
| 517 | return success(); |
| 518 | } |
| 519 | return failure(); |
| 520 | } |
| 521 | |
| 522 | private: |
| 523 | DataFlowSolver &solver; |
| 524 | SmallVector<unsigned, 4> targetBitwidths; |
| 525 | }; |
| 526 | |
| 527 | /// Fold index_cast(index_cast(%arg: i8, index), i8) -> %arg |
| 528 | /// This pattern assumes all passed `targetBitwidths` are not wider than index |
| 529 | /// type. |
| 530 | template <typename CastOp> |
| 531 | struct FoldIndexCastChain final : OpRewritePattern<CastOp> { |
| 532 | FoldIndexCastChain(MLIRContext *context, ArrayRef<unsigned> target) |
| 533 | : OpRewritePattern<CastOp>(context), targetBitwidths(target) {} |
| 534 | |
| 535 | LogicalResult matchAndRewrite(CastOp op, |
| 536 | PatternRewriter &rewriter) const override { |
| 537 | auto srcOp = op.getIn().template getDefiningOp<CastOp>(); |
| 538 | if (!srcOp) |
| 539 | return rewriter.notifyMatchFailure(op, "doesn't come from an index cast" ); |
| 540 | |
| 541 | Value src = srcOp.getIn(); |
| 542 | if (src.getType() != op.getType()) |
| 543 | return rewriter.notifyMatchFailure(op, "outer types don't match" ); |
| 544 | |
| 545 | if (!srcOp.getType().isIndex()) |
| 546 | return rewriter.notifyMatchFailure(op, "intermediate type isn't index" ); |
| 547 | |
| 548 | auto intType = dyn_cast<IntegerType>(op.getType()); |
| 549 | if (!intType || !llvm::is_contained(targetBitwidths, intType.getWidth())) |
| 550 | return failure(); |
| 551 | |
| 552 | rewriter.replaceOp(op, src); |
| 553 | return success(); |
| 554 | } |
| 555 | |
| 556 | private: |
| 557 | SmallVector<unsigned, 4> targetBitwidths; |
| 558 | }; |
| 559 | |
| 560 | struct NarrowLoopBounds final : OpInterfaceRewritePattern<LoopLikeOpInterface> { |
| 561 | NarrowLoopBounds(MLIRContext *context, DataFlowSolver &s, |
| 562 | ArrayRef<unsigned> target) |
| 563 | : OpInterfaceRewritePattern<LoopLikeOpInterface>(context), solver(s), |
| 564 | targetBitwidths(target), |
| 565 | boundsNarrowingFailedAttr( |
| 566 | StringAttr::get(context, bytes: "arith.bounds_narrowing_failed" )) {} |
| 567 | |
| 568 | LogicalResult matchAndRewrite(LoopLikeOpInterface loopLike, |
| 569 | PatternRewriter &rewriter) const override { |
| 570 | // Skip ops where bounds narrowing previously failed. |
| 571 | if (loopLike->hasDiscardableAttr(name: boundsNarrowingFailedAttr)) |
| 572 | return rewriter.notifyMatchFailure(arg&: loopLike, |
| 573 | msg: "bounds narrowing previously failed" ); |
| 574 | |
| 575 | std::optional<SmallVector<Value>> inductionVars = |
| 576 | loopLike.getLoopInductionVars(); |
| 577 | if (!inductionVars.has_value() || inductionVars->empty()) |
| 578 | return rewriter.notifyMatchFailure(arg&: loopLike, msg: "no induction variables" ); |
| 579 | |
| 580 | std::optional<SmallVector<OpFoldResult>> lowerBounds = |
| 581 | loopLike.getLoopLowerBounds(); |
| 582 | std::optional<SmallVector<OpFoldResult>> upperBounds = |
| 583 | loopLike.getLoopUpperBounds(); |
| 584 | std::optional<SmallVector<OpFoldResult>> steps = loopLike.getLoopSteps(); |
| 585 | |
| 586 | if (!lowerBounds.has_value() || !upperBounds.has_value() || |
| 587 | !steps.has_value()) |
| 588 | return rewriter.notifyMatchFailure(arg&: loopLike, msg: "no loop bounds or steps" ); |
| 589 | |
| 590 | if (lowerBounds->size() != inductionVars->size() || |
| 591 | upperBounds->size() != inductionVars->size() || |
| 592 | steps->size() != inductionVars->size()) |
| 593 | return rewriter.notifyMatchFailure(arg&: loopLike, |
| 594 | msg: "mismatched bounds/steps count" ); |
| 595 | |
| 596 | Location loc = loopLike->getLoc(); |
| 597 | SmallVector<OpFoldResult> newLowerBounds(*lowerBounds); |
| 598 | SmallVector<OpFoldResult> newUpperBounds(*upperBounds); |
| 599 | SmallVector<OpFoldResult> newSteps(*steps); |
| 600 | SmallVector<std::tuple<size_t, Type, CastKind>> narrowings; |
| 601 | |
| 602 | // Check each (indVar, lb, ub, step) tuple. |
| 603 | for (auto [idx, indVar, lbOFR, ubOFR, stepOFR] : |
| 604 | llvm::enumerate(First&: *inductionVars, Rest&: *lowerBounds, Rest&: *upperBounds, Rest&: *steps)) { |
| 605 | |
| 606 | // Only process value operands, skip attributes. |
| 607 | auto maybeLb = dyn_cast<Value>(Val&: lbOFR); |
| 608 | auto maybeUb = dyn_cast<Value>(Val&: ubOFR); |
| 609 | auto maybeStep = dyn_cast<Value>(Val&: stepOFR); |
| 610 | |
| 611 | if (!maybeLb || !maybeUb || !maybeStep) |
| 612 | continue; |
| 613 | |
| 614 | // Collect ranges for (lb, ub, step, indVar). |
| 615 | SmallVector<ConstantIntRanges> ranges; |
| 616 | if (failed(Result: collectRanges( |
| 617 | solver, values: ValueRange{maybeLb, maybeUb, maybeStep, indVar}, ranges))) |
| 618 | continue; |
| 619 | |
| 620 | const ConstantIntRanges &stepRange = ranges[2]; |
| 621 | const ConstantIntRanges &indVarRange = ranges[3]; |
| 622 | |
| 623 | Type srcType = maybeLb.getType(); |
| 624 | |
| 625 | // Try each target bitwidth. |
| 626 | for (unsigned targetBitwidth : targetBitwidths) { |
| 627 | Type targetType = getTargetType(srcType, targetBitwidth); |
| 628 | if (targetType == srcType) |
| 629 | continue; |
| 630 | |
| 631 | // Check if the target type is valid for this loop's induction |
| 632 | // variables. |
| 633 | if (!loopLike.isValidInductionVarType(type: targetType)) |
| 634 | continue; |
| 635 | |
| 636 | // Check if all values in this tuple can be truncated. |
| 637 | CastKind castKind = CastKind::Both; |
| 638 | for (const ConstantIntRanges &range : ranges) { |
| 639 | castKind = mergeCastKinds(lhs: castKind, |
| 640 | rhs: checkTruncatability(range, targetWidth: targetBitwidth)); |
| 641 | if (castKind == CastKind::None) |
| 642 | break; |
| 643 | } |
| 644 | |
| 645 | if (castKind == CastKind::None) |
| 646 | continue; |
| 647 | |
| 648 | // Check if indVar + step fits in the narrowed type. |
| 649 | // This is critical for loop correctness: the loop computes |
| 650 | // iv_next = iv_current + step in the narrowed type, then compares |
| 651 | // iv_next < ub. If iv_current + step overflows, the comparison may |
| 652 | // produce incorrect results and break loop termination. |
| 653 | // Both signed and unsigned interpretations must fit because loop |
| 654 | // semantics are unknown (integer types are signless). |
| 655 | ConstantIntRanges indVarPlusStepRange( |
| 656 | indVarRange.smin().sadd_sat(RHS: stepRange.smin()), |
| 657 | indVarRange.smax().sadd_sat(RHS: stepRange.smax()), |
| 658 | indVarRange.umin().uadd_sat(RHS: stepRange.umin()), |
| 659 | indVarRange.umax().uadd_sat(RHS: stepRange.umax())); |
| 660 | |
| 661 | if (checkTruncatability(range: indVarPlusStepRange, targetWidth: targetBitwidth) != |
| 662 | CastKind::Both) |
| 663 | continue; |
| 664 | |
| 665 | // Narrow the bounds and step values. |
| 666 | Value newLb = doCast(builder&: rewriter, loc, src: maybeLb, dstType: targetType, castKind); |
| 667 | Value newUb = doCast(builder&: rewriter, loc, src: maybeUb, dstType: targetType, castKind); |
| 668 | Value newStep = doCast(builder&: rewriter, loc, src: maybeStep, dstType: targetType, castKind); |
| 669 | |
| 670 | newLowerBounds[idx] = newLb; |
| 671 | newUpperBounds[idx] = newUb; |
| 672 | newSteps[idx] = newStep; |
| 673 | narrowings.push_back(Elt: {idx, targetType, castKind}); |
| 674 | break; |
| 675 | } |
| 676 | } |
| 677 | |
| 678 | if (narrowings.empty()) |
| 679 | return rewriter.notifyMatchFailure(arg&: loopLike, msg: "no narrowings found" ); |
| 680 | |
| 681 | // Save original types before modifying. |
| 682 | SmallVector<Type> origTypes; |
| 683 | for (auto [idx, targetType, castKind] : narrowings) { |
| 684 | Value indVar = (*inductionVars)[idx]; |
| 685 | origTypes.push_back(Elt: indVar.getType()); |
| 686 | } |
| 687 | |
| 688 | // Attempt to update bounds and induction variable types. |
| 689 | // If this fails, mark the op so we don't try again. |
| 690 | bool updateFailed = false; |
| 691 | rewriter.modifyOpInPlace(root: loopLike, callable: [&]() { |
| 692 | // Update the loop bounds and steps. |
| 693 | if (failed(Result: loopLike.setLoopLowerBounds(newLowerBounds)) || |
| 694 | failed(Result: loopLike.setLoopUpperBounds(newUpperBounds)) || |
| 695 | failed(Result: loopLike.setLoopSteps(newSteps))) { |
| 696 | // Mark op to prevent future attempts. IR was modified (attribute |
| 697 | // added), so we must return success() from the pattern. |
| 698 | loopLike->setDiscardableAttr(name: boundsNarrowingFailedAttr, |
| 699 | value: rewriter.getUnitAttr()); |
| 700 | updateFailed = true; |
| 701 | return; |
| 702 | } |
| 703 | |
| 704 | // Update induction variable types. |
| 705 | for (auto [idx, targetType, castKind] : narrowings) { |
| 706 | Value indVar = (*inductionVars)[idx]; |
| 707 | auto blockArg = cast<BlockArgument>(Val&: indVar); |
| 708 | |
| 709 | // Change the block argument type. |
| 710 | blockArg.setType(targetType); |
| 711 | } |
| 712 | }); |
| 713 | |
| 714 | if (updateFailed) |
| 715 | return success(); |
| 716 | |
| 717 | // Insert casts back to original type for uses. |
| 718 | for (auto [narrowingIdx, narrowingInfo] : llvm::enumerate(First&: narrowings)) { |
| 719 | auto [idx, targetType, castKind] = narrowingInfo; |
| 720 | Value indVar = (*inductionVars)[idx]; |
| 721 | auto blockArg = cast<BlockArgument>(Val&: indVar); |
| 722 | Type origType = origTypes[narrowingIdx]; |
| 723 | |
| 724 | OpBuilder::InsertionGuard guard(rewriter); |
| 725 | rewriter.setInsertionPointToStart(blockArg.getOwner()); |
| 726 | Value casted = doCast(builder&: rewriter, loc, src: blockArg, dstType: origType, castKind); |
| 727 | copyIntegerRange(solver, oldVal: blockArg, newVal: casted); |
| 728 | |
| 729 | // Replace all uses of the narrowed indVar with the casted value. |
| 730 | rewriter.replaceAllUsesExcept(from: blockArg, to: casted, exceptedUser: casted.getDefiningOp()); |
| 731 | } |
| 732 | |
| 733 | return success(); |
| 734 | } |
| 735 | |
| 736 | private: |
| 737 | DataFlowSolver &solver; |
| 738 | SmallVector<unsigned, 4> targetBitwidths; |
| 739 | StringAttr boundsNarrowingFailedAttr; |
| 740 | }; |
| 741 | |
| 742 | struct IntRangeOptimizationsPass final |
| 743 | : arith::impl::ArithIntRangeOptsBase<IntRangeOptimizationsPass> { |
| 744 | |
| 745 | void runOnOperation() override { |
| 746 | Operation *op = getOperation(); |
| 747 | MLIRContext *ctx = op->getContext(); |
| 748 | DataFlowSolver solver; |
| 749 | loadBaselineAnalyses(solver); |
| 750 | solver.load<IntegerRangeAnalysis>(); |
| 751 | if (failed(Result: solver.initializeAndRun(top: op))) |
| 752 | return signalPassFailure(); |
| 753 | |
| 754 | DataFlowListener listener(solver); |
| 755 | |
| 756 | RewritePatternSet patterns(ctx); |
| 757 | populateIntRangeOptimizationsPatterns(patterns, solver); |
| 758 | |
| 759 | // Disable folding and region simplification to avoid breaking the solver |
| 760 | // state. Both can remove block arguments (folding via control-flow |
| 761 | // simplification, region simplification via dead-arg elimination), which |
| 762 | // frees their underlying storage. A subsequent allocation may reuse the |
| 763 | // same address for a different block argument, causing stale solver state |
| 764 | // to be associated with the new argument and producing incorrect constants. |
| 765 | if (failed( |
| 766 | Result: applyPatternsGreedily(op, patterns: std::move(patterns), |
| 767 | config: GreedyRewriteConfig() |
| 768 | .enableFolding(enable: false) |
| 769 | .setRegionSimplificationLevel( |
| 770 | GreedySimplifyRegionLevel::Disabled) |
| 771 | .setListener(&listener)))) |
| 772 | signalPassFailure(); |
| 773 | } |
| 774 | }; |
| 775 | |
| 776 | struct IntRangeNarrowingPass final |
| 777 | : arith::impl::ArithIntRangeNarrowingBase<IntRangeNarrowingPass> { |
| 778 | using ArithIntRangeNarrowingBase::ArithIntRangeNarrowingBase; |
| 779 | |
| 780 | void runOnOperation() override { |
| 781 | Operation *op = getOperation(); |
| 782 | MLIRContext *ctx = op->getContext(); |
| 783 | DataFlowSolver solver; |
| 784 | loadBaselineAnalyses(solver); |
| 785 | solver.load<IntegerRangeAnalysis>(); |
| 786 | if (failed(Result: solver.initializeAndRun(top: op))) |
| 787 | return signalPassFailure(); |
| 788 | |
| 789 | DataFlowListener listener(solver); |
| 790 | |
| 791 | RewritePatternSet patterns(ctx); |
| 792 | populateIntRangeNarrowingPatterns(patterns, solver, bitwidthsSupported); |
| 793 | populateControlFlowValuesNarrowingPatterns(patterns, solver, |
| 794 | bitwidthsSupported); |
| 795 | |
| 796 | // We specifically need bottom-up traversal as cmpi pattern needs range |
| 797 | // data, attached to its original argument values. |
| 798 | if (failed(Result: applyPatternsGreedily( |
| 799 | op, patterns: std::move(patterns), |
| 800 | config: GreedyRewriteConfig().setUseTopDownTraversal(false).setListener( |
| 801 | &listener)))) |
| 802 | signalPassFailure(); |
| 803 | } |
| 804 | }; |
| 805 | } // namespace |
| 806 | |
| 807 | void mlir::arith::populateIntRangeOptimizationsPatterns( |
| 808 | RewritePatternSet &patterns, DataFlowSolver &solver) { |
| 809 | patterns.add<MaterializeKnownConstantValues, DeleteTrivialRem<RemSIOp>, |
| 810 | DeleteTrivialRem<RemUIOp>>(arg: patterns.getContext(), args&: solver); |
| 811 | } |
| 812 | |
| 813 | void mlir::arith::populateIntRangeNarrowingPatterns( |
| 814 | RewritePatternSet &patterns, DataFlowSolver &solver, |
| 815 | ArrayRef<unsigned> bitwidthsSupported) { |
| 816 | patterns.add<NarrowElementwise, NarrowCmpI>(arg: patterns.getContext(), args&: solver, |
| 817 | args&: bitwidthsSupported); |
| 818 | patterns.add<FoldIndexCastChain<arith::IndexCastUIOp>, |
| 819 | FoldIndexCastChain<arith::IndexCastOp>>(arg: patterns.getContext(), |
| 820 | args&: bitwidthsSupported); |
| 821 | } |
| 822 | |
| 823 | void mlir::arith::populateControlFlowValuesNarrowingPatterns( |
| 824 | RewritePatternSet &patterns, DataFlowSolver &solver, |
| 825 | ArrayRef<unsigned> bitwidthsSupported) { |
| 826 | patterns.add<NarrowLoopBounds>(arg: patterns.getContext(), args&: solver, |
| 827 | args&: bitwidthsSupported); |
| 828 | } |
| 829 | |
| 830 | std::unique_ptr<Pass> mlir::arith::createIntRangeOptimizationsPass() { |
| 831 | return std::make_unique<IntRangeOptimizationsPass>(); |
| 832 | } |
| 833 | |