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
isMemRefDereferencingOp(Operation & op)64 static bool isMemRefDereferencingOp(Operation &op) {
65 // TODO: Support DMA Ops.
66 return isa<AffineReadOpInterface, AffineWriteOpInterface>(op);
67 }
68
69 // Returns true if the individual op is loop invariant.
isOpLoopInvariant(Operation & op,Value indVar,SmallPtrSetImpl<Operation * > & definedOps,SmallPtrSetImpl<Operation * > & opsToHoist)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 // Register op in the set of ops defined inside the loop. This set is used
88 // to prevent hoisting ops that depend on other ops defined inside the loop
89 // which are themselves not being hoisted.
90 definedOps.insert(&op);
91
92 if (isMemRefDereferencingOp(op)) {
93 Value memref = isa<AffineReadOpInterface>(op)
94 ? cast<AffineReadOpInterface>(op).getMemRef()
95 : cast<AffineWriteOpInterface>(op).getMemRef();
96 for (auto *user : memref.getUsers()) {
97 // If this memref has a user that is a DMA, give up because these
98 // operations write to this memref.
99 if (isa<AffineDmaStartOp, AffineDmaWaitOp>(op)) {
100 return false;
101 }
102 // If the memref used by the load/store is used in a store elsewhere in
103 // the loop nest, we do not hoist. Similarly, if the memref used in a
104 // load is also being stored too, we do not hoist the load.
105 if (isa<AffineWriteOpInterface>(user) ||
106 (isa<AffineReadOpInterface>(user) &&
107 isa<AffineWriteOpInterface>(op))) {
108 if (&op != user) {
109 SmallVector<AffineForOp, 8> userIVs;
110 getLoopIVs(*user, &userIVs);
111 // Check that userIVs don't contain the for loop around the op.
112 if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) {
113 return false;
114 }
115 }
116 }
117 }
118 }
119
120 if (op.getNumOperands() == 0 && !isa<AffineYieldOp>(op)) {
121 LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n");
122 return false;
123 }
124 for (unsigned int i = 0; i < op.getNumOperands(); ++i) {
125 auto *operandSrc = op.getOperand(i).getDefiningOp();
126
127 LLVM_DEBUG(
128 op.getOperand(i).print(llvm::dbgs() << "\nIterating on operand\n"));
129
130 // If the loop IV is the operand, this op isn't loop invariant.
131 if (indVar == op.getOperand(i)) {
132 LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n");
133 return false;
134 }
135
136 if (operandSrc != nullptr) {
137 LLVM_DEBUG(llvm::dbgs()
138 << *operandSrc << "\nIterating on operand src\n");
139
140 // If the value was defined in the loop (outside of the
141 // if/else region), and that operation itself wasn't meant to
142 // be hoisted, then mark this operation loop dependent.
143 if (definedOps.count(operandSrc) && opsToHoist.count(operandSrc) == 0) {
144 return false;
145 }
146 }
147 }
148 }
149
150 // If no operand was loop variant, mark this op for motion.
151 opsToHoist.insert(&op);
152 return true;
153 }
154
155 // Checks if all ops in a region (i.e. list of blocks) are loop invariant.
areAllOpsInTheBlockListInvariant(Region & blockList,Value indVar,SmallPtrSetImpl<Operation * > & definedOps,SmallPtrSetImpl<Operation * > & opsToHoist)156 bool areAllOpsInTheBlockListInvariant(
157 Region &blockList, Value indVar, SmallPtrSetImpl<Operation *> &definedOps,
158 SmallPtrSetImpl<Operation *> &opsToHoist) {
159
160 for (auto &b : blockList) {
161 for (auto &op : b) {
162 if (!isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
163 return false;
164 }
165 }
166 }
167
168 return true;
169 }
170
171 // Returns true if the affine.if op can be hoisted.
checkInvarianceOfNestedIfOps(Operation * op,Value indVar,SmallPtrSetImpl<Operation * > & definedOps,SmallPtrSetImpl<Operation * > & opsToHoist)172 bool checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
173 SmallPtrSetImpl<Operation *> &definedOps,
174 SmallPtrSetImpl<Operation *> &opsToHoist) {
175 assert(isa<AffineIfOp>(op));
176 auto ifOp = cast<AffineIfOp>(op);
177
178 if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, definedOps,
179 opsToHoist)) {
180 return false;
181 }
182
183 if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, definedOps,
184 opsToHoist)) {
185 return false;
186 }
187
188 return true;
189 }
190
runOnAffineForOp(AffineForOp forOp)191 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) {
192 auto *loopBody = forOp.getBody();
193 auto indVar = forOp.getInductionVar();
194
195 SmallPtrSet<Operation *, 8> definedOps;
196 // This is the place where hoisted instructions would reside.
197 OpBuilder b(forOp.getOperation());
198
199 SmallPtrSet<Operation *, 8> opsToHoist;
200 SmallVector<Operation *, 8> opsToMove;
201
202 for (auto &op : *loopBody) {
203 // We don't hoist for loops.
204 if (!isa<AffineForOp>(op)) {
205 if (!isa<AffineYieldOp>(op)) {
206 if (isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
207 opsToMove.push_back(&op);
208 }
209 }
210 }
211 }
212
213 // For all instructions that we found to be invariant, place sequentially
214 // right before the for loop.
215 for (auto *op : opsToMove) {
216 op->moveBefore(forOp);
217 }
218
219 LLVM_DEBUG(forOp->print(llvm::dbgs() << "Modified loop\n"));
220 }
221
runOnFunction()222 void LoopInvariantCodeMotion::runOnFunction() {
223 // Walk through all loops in a function in innermost-loop-first order. This
224 // way, we first LICM from the inner loop, and place the ops in
225 // the outer loop, which in turn can be further LICM'ed.
226 getFunction().walk([&](AffineForOp op) {
227 LLVM_DEBUG(op->print(llvm::dbgs() << "\nOriginal loop\n"));
228 runOnAffineForOp(op);
229 });
230 }
231
232 std::unique_ptr<OperationPass<FuncOp>>
createAffineLoopInvariantCodeMotionPass()233 mlir::createAffineLoopInvariantCodeMotionPass() {
234 return std::make_unique<LoopInvariantCodeMotion>();
235 }
236