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 "PassDetail.h"
14 #include "mlir/Analysis/AffineAnalysis.h"
15 #include "mlir/Analysis/AffineStructures.h"
16 #include "mlir/Analysis/LoopAnalysis.h"
17 #include "mlir/Analysis/SliceAnalysis.h"
18 #include "mlir/Analysis/Utils.h"
19 #include "mlir/Dialect/Affine/IR/AffineOps.h"
20 #include "mlir/Dialect/Affine/Passes.h"
21 #include "mlir/IR/AffineExpr.h"
22 #include "mlir/IR/AffineMap.h"
23 #include "mlir/IR/Builders.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: The pass is missing zero-trip tests.
41 /// TODO: 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
45     : public AffineLoopInvariantCodeMotionBase<LoopInvariantCodeMotion> {
46   void runOnFunction() override;
47   void runOnAffineForOp(AffineForOp forOp);
48 };
49 } // end anonymous namespace
50 
51 static bool
52 checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
53                              SmallPtrSetImpl<Operation *> &opsWithUsers,
54                              SmallPtrSetImpl<Operation *> &opsToHoist);
55 static bool isOpLoopInvariant(Operation &op, Value indVar,
56                               SmallPtrSetImpl<Operation *> &opsWithUsers,
57                               SmallPtrSetImpl<Operation *> &opsToHoist);
58 
59 static bool
60 areAllOpsInTheBlockListInvariant(Region &blockList, Value indVar,
61                                  SmallPtrSetImpl<Operation *> &opsWithUsers,
62                                  SmallPtrSetImpl<Operation *> &opsToHoist);
63 
64 // Returns true if the individual op is loop invariant.
65 bool isOpLoopInvariant(Operation &op, Value indVar,
66                        SmallPtrSetImpl<Operation *> &opsWithUsers,
67                        SmallPtrSetImpl<Operation *> &opsToHoist) {
68   LLVM_DEBUG(llvm::dbgs() << "iterating on op: " << op;);
69 
70   if (isa<AffineIfOp>(op)) {
71     if (!checkInvarianceOfNestedIfOps(&op, indVar, opsWithUsers, opsToHoist)) {
72       return false;
73     }
74   } else if (auto forOp = dyn_cast<AffineForOp>(op)) {
75     if (!areAllOpsInTheBlockListInvariant(forOp.getLoopBody(), indVar,
76                                           opsWithUsers, opsToHoist)) {
77       return false;
78     }
79   } else if (isa<AffineDmaStartOp, AffineDmaWaitOp>(op)) {
80     // TODO: Support DMA ops.
81     return false;
82   } else if (!isa<ConstantOp>(op)) {
83     // Register op in the set of ops that have users.
84     opsWithUsers.insert(&op);
85     if (isa<AffineMapAccessInterface>(op)) {
86       Value memref = isa<AffineReadOpInterface>(op)
87                          ? cast<AffineReadOpInterface>(op).getMemRef()
88                          : cast<AffineWriteOpInterface>(op).getMemRef();
89       for (auto *user : memref.getUsers()) {
90         // If this memref has a user that is a DMA, give up because these
91         // operations write to this memref.
92         if (isa<AffineDmaStartOp, AffineDmaWaitOp>(op)) {
93           return false;
94         }
95         // If the memref used by the load/store is used in a store elsewhere in
96         // the loop nest, we do not hoist. Similarly, if the memref used in a
97         // load is also being stored too, we do not hoist the load.
98         if (isa<AffineWriteOpInterface>(user) ||
99             (isa<AffineReadOpInterface>(user) &&
100              isa<AffineWriteOpInterface>(op))) {
101           if (&op != user) {
102             SmallVector<AffineForOp, 8> userIVs;
103             getLoopIVs(*user, &userIVs);
104             // Check that userIVs don't contain the for loop around the op.
105             if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) {
106               return false;
107             }
108           }
109         }
110       }
111     }
112 
113     if (op.getNumOperands() == 0 && !isa<AffineYieldOp>(op)) {
114       LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n");
115       return false;
116     }
117   }
118 
119   // Check operands.
120   for (unsigned int i = 0; i < op.getNumOperands(); ++i) {
121     auto *operandSrc = op.getOperand(i).getDefiningOp();
122 
123     LLVM_DEBUG(
124         op.getOperand(i).print(llvm::dbgs() << "\nIterating on operand\n"));
125 
126     // If the loop IV is the operand, this op isn't loop invariant.
127     if (indVar == op.getOperand(i)) {
128       LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n");
129       return false;
130     }
131 
132     if (operandSrc != nullptr) {
133       LLVM_DEBUG(llvm::dbgs() << *operandSrc << "\nIterating on operand src\n");
134 
135       // If the value was defined in the loop (outside of the
136       // if/else region), and that operation itself wasn't meant to
137       // be hoisted, then mark this operation loop dependent.
138       if (opsWithUsers.count(operandSrc) && opsToHoist.count(operandSrc) == 0) {
139         return false;
140       }
141     }
142   }
143 
144   // If no operand was loop variant, mark this op for motion.
145   opsToHoist.insert(&op);
146   return true;
147 }
148 
149 // Checks if all ops in a region (i.e. list of blocks) are loop invariant.
150 bool areAllOpsInTheBlockListInvariant(
151     Region &blockList, Value indVar, SmallPtrSetImpl<Operation *> &opsWithUsers,
152     SmallPtrSetImpl<Operation *> &opsToHoist) {
153 
154   for (auto &b : blockList) {
155     for (auto &op : b) {
156       if (!isOpLoopInvariant(op, indVar, opsWithUsers, opsToHoist)) {
157         return false;
158       }
159     }
160   }
161 
162   return true;
163 }
164 
165 // Returns true if the affine.if op can be hoisted.
166 bool checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
167                                   SmallPtrSetImpl<Operation *> &opsWithUsers,
168                                   SmallPtrSetImpl<Operation *> &opsToHoist) {
169   assert(isa<AffineIfOp>(op));
170   auto ifOp = cast<AffineIfOp>(op);
171 
172   if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, opsWithUsers,
173                                         opsToHoist)) {
174     return false;
175   }
176 
177   if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, opsWithUsers,
178                                         opsToHoist)) {
179     return false;
180   }
181 
182   return true;
183 }
184 
185 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) {
186   auto *loopBody = forOp.getBody();
187   auto indVar = forOp.getInductionVar();
188 
189   // This is the place where hoisted instructions would reside.
190   OpBuilder b(forOp.getOperation());
191 
192   SmallPtrSet<Operation *, 8> opsToHoist;
193   SmallVector<Operation *, 8> opsToMove;
194   SmallPtrSet<Operation *, 8> opsWithUsers;
195 
196   for (auto &op : *loopBody) {
197     // Register op in the set of ops that have users. This set is used
198     // to prevent hoisting ops that depend on these ops that are
199     // not being hoisted.
200     if (!op.use_empty())
201       opsWithUsers.insert(&op);
202     if (!isa<AffineYieldOp>(op)) {
203       if (isOpLoopInvariant(op, indVar, opsWithUsers, opsToHoist)) {
204         opsToMove.push_back(&op);
205       }
206     }
207   }
208 
209   // For all instructions that we found to be invariant, place sequentially
210   // right before the for loop.
211   for (auto *op : opsToMove) {
212     op->moveBefore(forOp);
213   }
214 
215   LLVM_DEBUG(forOp->print(llvm::dbgs() << "Modified loop\n"));
216 }
217 
218 void LoopInvariantCodeMotion::runOnFunction() {
219   // Walk through all loops in a function in innermost-loop-first order.  This
220   // way, we first LICM from the inner loop, and place the ops in
221   // the outer loop, which in turn can be further LICM'ed.
222   getFunction().walk([&](AffineForOp op) {
223     LLVM_DEBUG(op->print(llvm::dbgs() << "\nOriginal loop\n"));
224     runOnAffineForOp(op);
225   });
226 }
227 
228 std::unique_ptr<OperationPass<FuncOp>>
229 mlir::createAffineLoopInvariantCodeMotionPass() {
230   return std::make_unique<LoopInvariantCodeMotion>();
231 }
232