1 //===- LoopInvariantCodeMotion.cpp - Code to perform loop fusion-----------===//
2 //
3 // Copyright 2019 The MLIR Authors.
4 //
5 // Licensed under the Apache License, Version 2.0 (the "License");
6 // you may not use this file except in compliance with the License.
7 // You may obtain a copy of the License at
8 //
9 //   http://www.apache.org/licenses/LICENSE-2.0
10 //
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS,
13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 // See the License for the specific language governing permissions and
15 // limitations under the License.
16 // =============================================================================
17 //
18 // This file implements loop invariant code motion.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "mlir/AffineOps/AffineOps.h"
23 #include "mlir/Analysis/AffineAnalysis.h"
24 #include "mlir/Analysis/AffineStructures.h"
25 #include "mlir/Analysis/LoopAnalysis.h"
26 #include "mlir/Analysis/SliceAnalysis.h"
27 #include "mlir/Analysis/Utils.h"
28 #include "mlir/IR/AffineExpr.h"
29 #include "mlir/IR/AffineMap.h"
30 #include "mlir/IR/Builders.h"
31 #include "mlir/Pass/Pass.h"
32 #include "mlir/StandardOps/Ops.h"
33 #include "mlir/Transforms/LoopUtils.h"
34 #include "mlir/Transforms/Passes.h"
35 #include "mlir/Transforms/Utils.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DenseSet.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 
43 #define DEBUG_TYPE "licm"
44 
45 using namespace mlir;
46 
47 namespace {
48 
49 /// Loop invariant code motion (LICM) pass.
50 /// TODO(asabne) : The pass is missing zero-trip tests.
51 /// TODO(asabne) : Check for the presence of side effects before hoisting.
52 struct LoopInvariantCodeMotion : public FunctionPass<LoopInvariantCodeMotion> {
53   void runOnFunction() override;
54   void runOnAffineForOp(AffineForOp forOp);
55 };
56 } // end anonymous namespace
57 
58 static bool
59 checkInvarianceOfNestedIfOps(Operation *op, Value *indVar,
60                              SmallPtrSetImpl<Operation *> &definedOps,
61                              SmallPtrSetImpl<Operation *> &opsToHoist);
62 static bool isOpLoopInvariant(Operation &op, Value *indVar,
63                               SmallPtrSetImpl<Operation *> &definedOps,
64                               SmallPtrSetImpl<Operation *> &opsToHoist);
65 
66 static bool
67 areAllOpsInTheBlockListInvariant(Region &blockList, Value *indVar,
68                                  SmallPtrSetImpl<Operation *> &definedOps,
69                                  SmallPtrSetImpl<Operation *> &opsToHoist);
70 
71 static bool isMemRefDereferencingOp(Operation &op) {
72   // TODO(asabne): Support DMA Ops.
73   if (isa<AffineLoadOp>(op) || isa<AffineStoreOp>(op)) {
74     return true;
75   }
76   return false;
77 }
78 
79 std::unique_ptr<FunctionPassBase> mlir::createLoopInvariantCodeMotionPass() {
80   return std::make_unique<LoopInvariantCodeMotion>();
81 }
82 
83 // Returns true if the individual op is loop invariant.
84 bool isOpLoopInvariant(Operation &op, Value *indVar,
85                        SmallPtrSetImpl<Operation *> &definedOps,
86                        SmallPtrSetImpl<Operation *> &opsToHoist) {
87   LLVM_DEBUG(llvm::dbgs() << "iterating on op: " << op;);
88 
89   if (isa<AffineIfOp>(op)) {
90     if (!checkInvarianceOfNestedIfOps(&op, indVar, definedOps, opsToHoist)) {
91       return false;
92     }
93   } else if (isa<AffineForOp>(op)) {
94     // If the body of a predicated region has a for loop, we don't hoist the
95     // 'affine.if'.
96     return false;
97   } else if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) {
98     // TODO(asabne): Support DMA ops.
99     return false;
100   } else if (!isa<ConstantOp>(op)) {
101     if (isMemRefDereferencingOp(op)) {
102       Value *memref = isa<AffineLoadOp>(op)
103                           ? cast<AffineLoadOp>(op).getMemRef()
104                           : cast<AffineStoreOp>(op).getMemRef();
105       for (auto *user : memref->getUsers()) {
106         // If this memref has a user that is a DMA, give up because these
107         // operations write to this memref.
108         if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) {
109           return false;
110         }
111         // If the memref used by the load/store is used in a store elsewhere in
112         // the loop nest, we do not hoist. Similarly, if the memref used in a
113         // load is also being stored too, we do not hoist the load.
114         if (isa<AffineStoreOp>(user) ||
115             (isa<AffineLoadOp>(user) && isa<AffineStoreOp>(op))) {
116           if (&op != user) {
117             SmallVector<AffineForOp, 8> userIVs;
118             getLoopIVs(*user, &userIVs);
119             // Check that userIVs don't contain the for loop around the op.
120             if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) {
121               return false;
122             }
123           }
124         }
125       }
126     }
127 
128     // Insert this op in the defined ops list.
129     definedOps.insert(&op);
130 
131     if (op.getNumOperands() == 0 && !isa<AffineTerminatorOp>(op)) {
132       LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n");
133       return false;
134     }
135     for (unsigned int i = 0; i < op.getNumOperands(); ++i) {
136       auto *operandSrc = op.getOperand(i)->getDefiningOp();
137 
138       LLVM_DEBUG(
139           op.getOperand(i)->print(llvm::dbgs() << "\nIterating on operand\n"));
140 
141       // If the loop IV is the operand, this op isn't loop invariant.
142       if (indVar == op.getOperand(i)) {
143         LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n");
144         return false;
145       }
146 
147       if (operandSrc != nullptr) {
148         LLVM_DEBUG(llvm::dbgs()
149                    << *operandSrc << "\nIterating on operand src\n");
150 
151         // If the value was defined in the loop (outside of the
152         // if/else region), and that operation itself wasn't meant to
153         // be hoisted, then mark this operation loop dependent.
154         if (definedOps.count(operandSrc) && opsToHoist.count(operandSrc) == 0) {
155           return false;
156         }
157       }
158     }
159   }
160 
161   // If no operand was loop variant, mark this op for motion.
162   opsToHoist.insert(&op);
163   return true;
164 }
165 
166 // Checks if all ops in a region (i.e. list of blocks) are loop invariant.
167 bool areAllOpsInTheBlockListInvariant(
168     Region &blockList, Value *indVar, SmallPtrSetImpl<Operation *> &definedOps,
169     SmallPtrSetImpl<Operation *> &opsToHoist) {
170 
171   for (auto &b : blockList) {
172     for (auto &op : b) {
173       if (!isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
174         return false;
175       }
176     }
177   }
178 
179   return true;
180 }
181 
182 // Returns true if the affine.if op can be hoisted.
183 bool checkInvarianceOfNestedIfOps(Operation *op, Value *indVar,
184                                   SmallPtrSetImpl<Operation *> &definedOps,
185                                   SmallPtrSetImpl<Operation *> &opsToHoist) {
186   assert(isa<AffineIfOp>(op));
187   auto ifOp = cast<AffineIfOp>(op);
188 
189   if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, definedOps,
190                                         opsToHoist)) {
191     return false;
192   }
193 
194   if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, definedOps,
195                                         opsToHoist)) {
196     return false;
197   }
198 
199   return true;
200 }
201 
202 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) {
203   auto *loopBody = forOp.getBody();
204   auto *indVar = forOp.getInductionVar();
205 
206   SmallPtrSet<Operation *, 8> definedOps;
207   // This is the place where hoisted instructions would reside.
208   OpBuilder b(forOp.getOperation());
209 
210   SmallPtrSet<Operation *, 8> opsToHoist;
211   SmallVector<Operation *, 8> opsToMove;
212 
213   for (auto &op : *loopBody) {
214     // We don't hoist for loops.
215     if (!isa<AffineForOp>(op)) {
216       if (!isa<AffineTerminatorOp>(op)) {
217         if (isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
218           opsToMove.push_back(&op);
219         }
220       }
221     }
222   }
223 
224   // For all instructions that we found to be invariant, place sequentially
225   // right before the for loop.
226   for (auto *op : opsToMove) {
227     op->moveBefore(forOp);
228   }
229 
230   LLVM_DEBUG(forOp.getOperation()->print(llvm::dbgs() << "Modified loop\n"));
231 
232   // If the for loop body has a single operation (the terminator), erase it.
233   if (forOp.getBody()->getOperations().size() == 1) {
234     assert(isa<AffineTerminatorOp>(forOp.getBody()->front()));
235     forOp.erase();
236   }
237 }
238 
239 void LoopInvariantCodeMotion::runOnFunction() {
240   // Walk through all loops in a function in innermost-loop-first order.  This
241   // way, we first LICM from the inner loop, and place the ops in
242   // the outer loop, which in turn can be further LICM'ed.
243   getFunction().walk<AffineForOp>([&](AffineForOp op) {
244     LLVM_DEBUG(op.getOperation()->print(llvm::dbgs() << "\nOriginal loop\n"));
245     runOnAffineForOp(op);
246   });
247 }
248 
249 static PassRegistration<LoopInvariantCodeMotion>
250     pass("affine-loop-invariant-code-motion",
251          "Hoist loop invariant instructions outside of the loop");
252