1 //===- AffineLoopInvariantCodeMotion.cpp - Code to perform loop fusion-----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements loop invariant code motion. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "mlir/Analysis/AffineAnalysis.h" 14 #include "mlir/Analysis/AffineStructures.h" 15 #include "mlir/Analysis/LoopAnalysis.h" 16 #include "mlir/Analysis/SliceAnalysis.h" 17 #include "mlir/Analysis/Utils.h" 18 #include "mlir/Dialect/Affine/IR/AffineOps.h" 19 #include "mlir/Dialect/Affine/Passes.h" 20 #include "mlir/IR/AffineExpr.h" 21 #include "mlir/IR/AffineMap.h" 22 #include "mlir/IR/Builders.h" 23 #include "mlir/Pass/Pass.h" 24 #include "mlir/Transforms/LoopUtils.h" 25 #include "mlir/Transforms/Utils.h" 26 #include "llvm/ADT/DenseMap.h" 27 #include "llvm/ADT/DenseSet.h" 28 #include "llvm/ADT/SmallPtrSet.h" 29 #include "llvm/Support/CommandLine.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/raw_ostream.h" 32 33 #define DEBUG_TYPE "licm" 34 35 using namespace mlir; 36 37 namespace { 38 39 /// Loop invariant code motion (LICM) pass. 40 /// TODO(asabne) : The pass is missing zero-trip tests. 41 /// TODO(asabne) : Check for the presence of side effects before hoisting. 42 /// TODO: This code should be removed once the new LICM pass can handle its 43 /// uses. 44 struct LoopInvariantCodeMotion : public FunctionPass<LoopInvariantCodeMotion> { 45 void runOnFunction() override; 46 void runOnAffineForOp(AffineForOp forOp); 47 }; 48 } // end anonymous namespace 49 50 static bool 51 checkInvarianceOfNestedIfOps(Operation *op, Value indVar, 52 SmallPtrSetImpl<Operation *> &definedOps, 53 SmallPtrSetImpl<Operation *> &opsToHoist); 54 static bool isOpLoopInvariant(Operation &op, Value indVar, 55 SmallPtrSetImpl<Operation *> &definedOps, 56 SmallPtrSetImpl<Operation *> &opsToHoist); 57 58 static bool 59 areAllOpsInTheBlockListInvariant(Region &blockList, Value indVar, 60 SmallPtrSetImpl<Operation *> &definedOps, 61 SmallPtrSetImpl<Operation *> &opsToHoist); 62 63 static bool isMemRefDereferencingOp(Operation &op) { 64 // TODO(asabne): Support DMA Ops. 65 if (isa<AffineLoadOp>(op) || isa<AffineStoreOp>(op)) { 66 return true; 67 } 68 return false; 69 } 70 71 // Returns true if the individual op is loop invariant. 72 bool isOpLoopInvariant(Operation &op, Value indVar, 73 SmallPtrSetImpl<Operation *> &definedOps, 74 SmallPtrSetImpl<Operation *> &opsToHoist) { 75 LLVM_DEBUG(llvm::dbgs() << "iterating on op: " << op;); 76 77 if (isa<AffineIfOp>(op)) { 78 if (!checkInvarianceOfNestedIfOps(&op, indVar, definedOps, opsToHoist)) { 79 return false; 80 } 81 } else if (isa<AffineForOp>(op)) { 82 // If the body of a predicated region has a for loop, we don't hoist the 83 // 'affine.if'. 84 return false; 85 } else if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) { 86 // TODO(asabne): Support DMA ops. 87 return false; 88 } else if (!isa<ConstantOp>(op)) { 89 if (isMemRefDereferencingOp(op)) { 90 Value memref = isa<AffineLoadOp>(op) 91 ? cast<AffineLoadOp>(op).getMemRef() 92 : cast<AffineStoreOp>(op).getMemRef(); 93 for (auto *user : memref.getUsers()) { 94 // If this memref has a user that is a DMA, give up because these 95 // operations write to this memref. 96 if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) { 97 return false; 98 } 99 // If the memref used by the load/store is used in a store elsewhere in 100 // the loop nest, we do not hoist. Similarly, if the memref used in a 101 // load is also being stored too, we do not hoist the load. 102 if (isa<AffineStoreOp>(user) || 103 (isa<AffineLoadOp>(user) && isa<AffineStoreOp>(op))) { 104 if (&op != user) { 105 SmallVector<AffineForOp, 8> userIVs; 106 getLoopIVs(*user, &userIVs); 107 // Check that userIVs don't contain the for loop around the op. 108 if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) { 109 return false; 110 } 111 } 112 } 113 } 114 } 115 116 // Insert this op in the defined ops list. 117 definedOps.insert(&op); 118 119 if (op.getNumOperands() == 0 && !isa<AffineTerminatorOp>(op)) { 120 LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n"); 121 return false; 122 } 123 for (unsigned int i = 0; i < op.getNumOperands(); ++i) { 124 auto *operandSrc = op.getOperand(i).getDefiningOp(); 125 126 LLVM_DEBUG( 127 op.getOperand(i).print(llvm::dbgs() << "\nIterating on operand\n")); 128 129 // If the loop IV is the operand, this op isn't loop invariant. 130 if (indVar == op.getOperand(i)) { 131 LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n"); 132 return false; 133 } 134 135 if (operandSrc != nullptr) { 136 LLVM_DEBUG(llvm::dbgs() 137 << *operandSrc << "\nIterating on operand src\n"); 138 139 // If the value was defined in the loop (outside of the 140 // if/else region), and that operation itself wasn't meant to 141 // be hoisted, then mark this operation loop dependent. 142 if (definedOps.count(operandSrc) && opsToHoist.count(operandSrc) == 0) { 143 return false; 144 } 145 } 146 } 147 } 148 149 // If no operand was loop variant, mark this op for motion. 150 opsToHoist.insert(&op); 151 return true; 152 } 153 154 // Checks if all ops in a region (i.e. list of blocks) are loop invariant. 155 bool areAllOpsInTheBlockListInvariant( 156 Region &blockList, Value indVar, SmallPtrSetImpl<Operation *> &definedOps, 157 SmallPtrSetImpl<Operation *> &opsToHoist) { 158 159 for (auto &b : blockList) { 160 for (auto &op : b) { 161 if (!isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) { 162 return false; 163 } 164 } 165 } 166 167 return true; 168 } 169 170 // Returns true if the affine.if op can be hoisted. 171 bool checkInvarianceOfNestedIfOps(Operation *op, Value indVar, 172 SmallPtrSetImpl<Operation *> &definedOps, 173 SmallPtrSetImpl<Operation *> &opsToHoist) { 174 assert(isa<AffineIfOp>(op)); 175 auto ifOp = cast<AffineIfOp>(op); 176 177 if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, definedOps, 178 opsToHoist)) { 179 return false; 180 } 181 182 if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, definedOps, 183 opsToHoist)) { 184 return false; 185 } 186 187 return true; 188 } 189 190 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) { 191 auto *loopBody = forOp.getBody(); 192 auto indVar = forOp.getInductionVar(); 193 194 SmallPtrSet<Operation *, 8> definedOps; 195 // This is the place where hoisted instructions would reside. 196 OpBuilder b(forOp.getOperation()); 197 198 SmallPtrSet<Operation *, 8> opsToHoist; 199 SmallVector<Operation *, 8> opsToMove; 200 201 for (auto &op : *loopBody) { 202 // We don't hoist for loops. 203 if (!isa<AffineForOp>(op)) { 204 if (!isa<AffineTerminatorOp>(op)) { 205 if (isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) { 206 opsToMove.push_back(&op); 207 } 208 } 209 } 210 } 211 212 // For all instructions that we found to be invariant, place sequentially 213 // right before the for loop. 214 for (auto *op : opsToMove) { 215 op->moveBefore(forOp); 216 } 217 218 LLVM_DEBUG(forOp.getOperation()->print(llvm::dbgs() << "Modified loop\n")); 219 } 220 221 void LoopInvariantCodeMotion::runOnFunction() { 222 // Walk through all loops in a function in innermost-loop-first order. This 223 // way, we first LICM from the inner loop, and place the ops in 224 // the outer loop, which in turn can be further LICM'ed. 225 getFunction().walk([&](AffineForOp op) { 226 LLVM_DEBUG(op.getOperation()->print(llvm::dbgs() << "\nOriginal loop\n")); 227 runOnAffineForOp(op); 228 }); 229 } 230 231 std::unique_ptr<OpPassBase<FuncOp>> 232 mlir::createAffineLoopInvariantCodeMotionPass() { 233 return std::make_unique<LoopInvariantCodeMotion>(); 234 } 235 236 static PassRegistration<LoopInvariantCodeMotion> 237 pass("affine-loop-invariant-code-motion", 238 "Hoist loop invariant instructions outside of the loop"); 239