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
45     : public PassWrapper<LoopInvariantCodeMotion, FunctionPass> {
46 /// Include the generated pass utilities.
47 #define GEN_PASS_AffineLoopInvariantCodeMotion
48 #include "mlir/Dialect/Affine/Passes.h.inc"
49 
50   void runOnFunction() override;
51   void runOnAffineForOp(AffineForOp forOp);
52 };
53 } // end anonymous namespace
54 
55 static bool
56 checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
57                              SmallPtrSetImpl<Operation *> &definedOps,
58                              SmallPtrSetImpl<Operation *> &opsToHoist);
59 static bool isOpLoopInvariant(Operation &op, Value indVar,
60                               SmallPtrSetImpl<Operation *> &definedOps,
61                               SmallPtrSetImpl<Operation *> &opsToHoist);
62 
63 static bool
64 areAllOpsInTheBlockListInvariant(Region &blockList, Value indVar,
65                                  SmallPtrSetImpl<Operation *> &definedOps,
66                                  SmallPtrSetImpl<Operation *> &opsToHoist);
67 
68 static bool isMemRefDereferencingOp(Operation &op) {
69   // TODO(asabne): Support DMA Ops.
70   if (isa<AffineLoadOp>(op) || isa<AffineStoreOp>(op)) {
71     return true;
72   }
73   return false;
74 }
75 
76 // Returns true if the individual op is loop invariant.
77 bool isOpLoopInvariant(Operation &op, Value indVar,
78                        SmallPtrSetImpl<Operation *> &definedOps,
79                        SmallPtrSetImpl<Operation *> &opsToHoist) {
80   LLVM_DEBUG(llvm::dbgs() << "iterating on op: " << op;);
81 
82   if (isa<AffineIfOp>(op)) {
83     if (!checkInvarianceOfNestedIfOps(&op, indVar, definedOps, opsToHoist)) {
84       return false;
85     }
86   } else if (isa<AffineForOp>(op)) {
87     // If the body of a predicated region has a for loop, we don't hoist the
88     // 'affine.if'.
89     return false;
90   } else if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) {
91     // TODO(asabne): Support DMA ops.
92     return false;
93   } else if (!isa<ConstantOp>(op)) {
94     if (isMemRefDereferencingOp(op)) {
95       Value memref = isa<AffineLoadOp>(op)
96                          ? cast<AffineLoadOp>(op).getMemRef()
97                          : cast<AffineStoreOp>(op).getMemRef();
98       for (auto *user : memref.getUsers()) {
99         // If this memref has a user that is a DMA, give up because these
100         // operations write to this memref.
101         if (isa<AffineDmaStartOp>(op) || isa<AffineDmaWaitOp>(op)) {
102           return false;
103         }
104         // If the memref used by the load/store is used in a store elsewhere in
105         // the loop nest, we do not hoist. Similarly, if the memref used in a
106         // load is also being stored too, we do not hoist the load.
107         if (isa<AffineStoreOp>(user) ||
108             (isa<AffineLoadOp>(user) && isa<AffineStoreOp>(op))) {
109           if (&op != user) {
110             SmallVector<AffineForOp, 8> userIVs;
111             getLoopIVs(*user, &userIVs);
112             // Check that userIVs don't contain the for loop around the op.
113             if (llvm::is_contained(userIVs, getForInductionVarOwner(indVar))) {
114               return false;
115             }
116           }
117         }
118       }
119     }
120 
121     // Insert this op in the defined ops list.
122     definedOps.insert(&op);
123 
124     if (op.getNumOperands() == 0 && !isa<AffineTerminatorOp>(op)) {
125       LLVM_DEBUG(llvm::dbgs() << "\nNon-constant op with 0 operands\n");
126       return false;
127     }
128     for (unsigned int i = 0; i < op.getNumOperands(); ++i) {
129       auto *operandSrc = op.getOperand(i).getDefiningOp();
130 
131       LLVM_DEBUG(
132           op.getOperand(i).print(llvm::dbgs() << "\nIterating on operand\n"));
133 
134       // If the loop IV is the operand, this op isn't loop invariant.
135       if (indVar == op.getOperand(i)) {
136         LLVM_DEBUG(llvm::dbgs() << "\nLoop IV is the operand\n");
137         return false;
138       }
139 
140       if (operandSrc != nullptr) {
141         LLVM_DEBUG(llvm::dbgs()
142                    << *operandSrc << "\nIterating on operand src\n");
143 
144         // If the value was defined in the loop (outside of the
145         // if/else region), and that operation itself wasn't meant to
146         // be hoisted, then mark this operation loop dependent.
147         if (definedOps.count(operandSrc) && opsToHoist.count(operandSrc) == 0) {
148           return false;
149         }
150       }
151     }
152   }
153 
154   // If no operand was loop variant, mark this op for motion.
155   opsToHoist.insert(&op);
156   return true;
157 }
158 
159 // Checks if all ops in a region (i.e. list of blocks) are loop invariant.
160 bool areAllOpsInTheBlockListInvariant(
161     Region &blockList, Value indVar, SmallPtrSetImpl<Operation *> &definedOps,
162     SmallPtrSetImpl<Operation *> &opsToHoist) {
163 
164   for (auto &b : blockList) {
165     for (auto &op : b) {
166       if (!isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
167         return false;
168       }
169     }
170   }
171 
172   return true;
173 }
174 
175 // Returns true if the affine.if op can be hoisted.
176 bool checkInvarianceOfNestedIfOps(Operation *op, Value indVar,
177                                   SmallPtrSetImpl<Operation *> &definedOps,
178                                   SmallPtrSetImpl<Operation *> &opsToHoist) {
179   assert(isa<AffineIfOp>(op));
180   auto ifOp = cast<AffineIfOp>(op);
181 
182   if (!areAllOpsInTheBlockListInvariant(ifOp.thenRegion(), indVar, definedOps,
183                                         opsToHoist)) {
184     return false;
185   }
186 
187   if (!areAllOpsInTheBlockListInvariant(ifOp.elseRegion(), indVar, definedOps,
188                                         opsToHoist)) {
189     return false;
190   }
191 
192   return true;
193 }
194 
195 void LoopInvariantCodeMotion::runOnAffineForOp(AffineForOp forOp) {
196   auto *loopBody = forOp.getBody();
197   auto indVar = forOp.getInductionVar();
198 
199   SmallPtrSet<Operation *, 8> definedOps;
200   // This is the place where hoisted instructions would reside.
201   OpBuilder b(forOp.getOperation());
202 
203   SmallPtrSet<Operation *, 8> opsToHoist;
204   SmallVector<Operation *, 8> opsToMove;
205 
206   for (auto &op : *loopBody) {
207     // We don't hoist for loops.
208     if (!isa<AffineForOp>(op)) {
209       if (!isa<AffineTerminatorOp>(op)) {
210         if (isOpLoopInvariant(op, indVar, definedOps, opsToHoist)) {
211           opsToMove.push_back(&op);
212         }
213       }
214     }
215   }
216 
217   // For all instructions that we found to be invariant, place sequentially
218   // right before the for loop.
219   for (auto *op : opsToMove) {
220     op->moveBefore(forOp);
221   }
222 
223   LLVM_DEBUG(forOp.getOperation()->print(llvm::dbgs() << "Modified loop\n"));
224 }
225 
226 void LoopInvariantCodeMotion::runOnFunction() {
227   // Walk through all loops in a function in innermost-loop-first order.  This
228   // way, we first LICM from the inner loop, and place the ops in
229   // the outer loop, which in turn can be further LICM'ed.
230   getFunction().walk([&](AffineForOp op) {
231     LLVM_DEBUG(op.getOperation()->print(llvm::dbgs() << "\nOriginal loop\n"));
232     runOnAffineForOp(op);
233   });
234 }
235 
236 std::unique_ptr<OperationPass<FuncOp>>
237 mlir::createAffineLoopInvariantCodeMotionPass() {
238   return std::make_unique<LoopInvariantCodeMotion>();
239 }
240