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 *> &definedOps,
54                              SmallPtrSetImpl<Operation *> &opsToHoist);
55 static bool isOpLoopInvariant(Operation &op, Value indVar,
56                               SmallPtrSetImpl<Operation *> &definedOps,
57                               SmallPtrSetImpl<Operation *> &opsToHoist);
58 
59 static bool
60 areAllOpsInTheBlockListInvariant(Region &blockList, Value indVar,
61                                  SmallPtrSetImpl<Operation *> &definedOps,
62                                  SmallPtrSetImpl<Operation *> &opsToHoist);
63 
64 static bool isMemRefDereferencingOp(Operation &op) {
65   // TODO: Support DMA Ops.
66   return isa<AffineLoadOp, AffineStoreOp>(op);
67 }
68 
69 // Returns true if the individual op is loop invariant.
70 bool isOpLoopInvariant(Operation &op, Value indVar,
71                        SmallPtrSetImpl<Operation *> &definedOps,
72                        SmallPtrSetImpl<Operation *> &opsToHoist) {
73   LLVM_DEBUG(llvm::dbgs() << "iterating on op: " << op;);
74 
75   if (isa<AffineIfOp>(op)) {
76     if (!checkInvarianceOfNestedIfOps(&op, indVar, definedOps, opsToHoist)) {
77       return false;
78     }
79   } else if (isa<AffineForOp>(op)) {
80     // If the body of a predicated region has a for loop, we don't hoist the
81     // 'affine.if'.
82     return false;
83   } else if (isa<AffineDmaStartOp, AffineDmaWaitOp>(op)) {
84     // TODO: Support DMA ops.
85     return false;
86   } else if (!isa<ConstantOp>(op)) {
87     if (isMemRefDereferencingOp(op)) {
88       Value memref = isa<AffineLoadOp>(op)
89                          ? cast<AffineLoadOp>(op).getMemRef()
90                          : cast<AffineStoreOp>(op).getMemRef();
91       for (auto *user : memref.getUsers()) {
92         // If this memref has a user that is a DMA, give up because these
93         // operations write to this memref.
94         if (isa<AffineDmaStartOp, AffineDmaWaitOp>(op)) {
95           return false;
96         }
97         // If the memref used by the load/store is used in a store elsewhere in
98         // the loop nest, we do not hoist. Similarly, if the memref used in a
99         // load is also being stored too, we do not hoist the load.
100         if (isa<AffineStoreOp>(user) ||
101             (isa<AffineLoadOp>(user) && isa<AffineStoreOp>(op))) {
102           if (&op != user) {
103             SmallVector<AffineForOp, 8> userIVs;
104             getLoopIVs(*user, &userIVs);
105             // Check that userIVs don't contain the for loop around the op.
106             if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) {
107               return false;
108             }
109           }
110         }
111       }
112     }
113 
114     // Insert this op in the defined ops list.
115     definedOps.insert(&op);
116 
117     if (op.getNumOperands() == 0 && !isa<AffineYieldOp>(op)) {
118       LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n");
119       return false;
120     }
121     for (unsigned int i = 0; i < op.getNumOperands(); ++i) {
122       auto *operandSrc = op.getOperand(i).getDefiningOp();
123 
124       LLVM_DEBUG(
125           op.getOperand(i).print(llvm::dbgs() << "\nIterating on operand\n"));
126 
127       // If the loop IV is the operand, this op isn't loop invariant.
128       if (indVar == op.getOperand(i)) {
129         LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n");
130         return false;
131       }
132 
133       if (operandSrc != nullptr) {
134         LLVM_DEBUG(llvm::dbgs()
135                    << *operandSrc << "\nIterating on operand src\n");
136 
137         // If the value was defined in the loop (outside of the
138         // if/else region), and that operation itself wasn't meant to
139         // be hoisted, then mark this operation loop dependent.
140         if (definedOps.count(operandSrc) && opsToHoist.count(operandSrc) == 0) {
141           return false;
142         }
143       }
144     }
145   }
146 
147   // If no operand was loop variant, mark this op for motion.
148   opsToHoist.insert(&op);
149   return true;
150 }
151 
152 // Checks if all ops in a region (i.e. list of blocks) are loop invariant.
153 bool areAllOpsInTheBlockListInvariant(
154     Region &blockList, Value indVar, SmallPtrSetImpl<Operation *> &definedOps,
155     SmallPtrSetImpl<Operation *> &opsToHoist) {
156 
157   for (auto &b : blockList) {
158     for (auto &op : b) {
159       if (!isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
160         return false;
161       }
162     }
163   }
164 
165   return true;
166 }
167 
168 // Returns true if the affine.if op can be hoisted.
169 bool checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
170                                   SmallPtrSetImpl<Operation *> &definedOps,
171                                   SmallPtrSetImpl<Operation *> &opsToHoist) {
172   assert(isa<AffineIfOp>(op));
173   auto ifOp = cast<AffineIfOp>(op);
174 
175   if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, definedOps,
176                                         opsToHoist)) {
177     return false;
178   }
179 
180   if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, definedOps,
181                                         opsToHoist)) {
182     return false;
183   }
184 
185   return true;
186 }
187 
188 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) {
189   auto *loopBody = forOp.getBody();
190   auto indVar = forOp.getInductionVar();
191 
192   SmallPtrSet<Operation *, 8> definedOps;
193   // This is the place where hoisted instructions would reside.
194   OpBuilder b(forOp.getOperation());
195 
196   SmallPtrSet<Operation *, 8> opsToHoist;
197   SmallVector<Operation *, 8> opsToMove;
198 
199   for (auto &op : *loopBody) {
200     // We don't hoist for loops.
201     if (!isa<AffineForOp>(op)) {
202       if (!isa<AffineYieldOp>(op)) {
203         if (isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
204           opsToMove.push_back(&op);
205         }
206       }
207     }
208   }
209 
210   // For all instructions that we found to be invariant, place sequentially
211   // right before the for loop.
212   for (auto *op : opsToMove) {
213     op->moveBefore(forOp);
214   }
215 
216   LLVM_DEBUG(forOp.getOperation()->print(llvm::dbgs() << "Modified loop\n"));
217 }
218 
219 void LoopInvariantCodeMotion::runOnFunction() {
220   // Walk through all loops in a function in innermost-loop-first order.  This
221   // way, we first LICM from the inner loop, and place the ops in
222   // the outer loop, which in turn can be further LICM'ed.
223   getFunction().walk([&](AffineForOp op) {
224     LLVM_DEBUG(op.getOperation()->print(llvm::dbgs() << "\nOriginal loop\n"));
225     runOnAffineForOp(op);
226   });
227 }
228 
229 std::unique_ptr<OperationPass<FuncOp>>
230 mlir::createAffineLoopInvariantCodeMotionPass() {
231   return std::make_unique<LoopInvariantCodeMotion>();
232 }
233