1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
  | 
//===- GreedyPatternRewriteDriver.cpp - A greedy rewriter -----------------===//
//
// Copyright 2019 The MLIR Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// =============================================================================
//
// This file implements mlir::applyPatternsGreedily.
//
//===----------------------------------------------------------------------===//
#include "mlir/IR/Builders.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/StandardOps/Ops.h"
#include "llvm/ADT/DenseMap.h"
using namespace mlir;
namespace {
/// This is a worklist-driven driver for the PatternMatcher, which repeatedly
/// applies the locally optimal patterns in a roughly "bottom up" way.
class GreedyPatternRewriteDriver : public PatternRewriter {
public:
  explicit GreedyPatternRewriteDriver(Function &fn,
                                      OwningRewritePatternList &&patterns)
      : PatternRewriter(fn.getContext()), matcher(std::move(patterns), *this),
        builder(&fn) {
    worklist.reserve(64);
    // Add all operations to the worklist.
    fn.walk([&](Operation *op) { addToWorklist(op); });
  }
  /// Perform the rewrites.
  void simplifyFunction();
  void addToWorklist(Operation *op) {
    // Check to see if the worklist already contains this op.
    if (worklistMap.count(op))
      return;
    worklistMap[op] = worklist.size();
    worklist.push_back(op);
  }
  Operation *popFromWorklist() {
    auto *op = worklist.back();
    worklist.pop_back();
    // This operation is no longer in the worklist, keep worklistMap up to date.
    if (op)
      worklistMap.erase(op);
    return op;
  }
  /// If the specified operation is in the worklist, remove it.  If not, this is
  /// a no-op.
  void removeFromWorklist(Operation *op) {
    auto it = worklistMap.find(op);
    if (it != worklistMap.end()) {
      assert(worklist[it->second] == op && "malformed worklist data structure");
      worklist[it->second] = nullptr;
    }
  }
  // These are hooks implemented for PatternRewriter.
protected:
  // Implement the hook for creating operations, and make sure that newly
  // created ops are added to the worklist for processing.
  Operation *createOperation(const OperationState &state) override {
    auto *result = builder.createOperation(state);
    addToWorklist(result);
    return result;
  }
  // If an operation is about to be removed, make sure it is not in our
  // worklist anymore because we'd get dangling references to it.
  void notifyOperationRemoved(Operation *op) override {
    addToWorklist(op->getOperands());
    removeFromWorklist(op);
  }
  // When the root of a pattern is about to be replaced, it can trigger
  // simplifications to its users - make sure to add them to the worklist
  // before the root is changed.
  void notifyRootReplaced(Operation *op) override {
    for (auto *result : op->getResults())
      // TODO: Add a result->getUsers() iterator.
      for (auto &user : result->getUses())
        addToWorklist(user.getOwner());
  }
private:
  // Look over the provided operands for any defining operations that should
  // be re-added to the worklist. This function should be called when an
  // operation is modified or removed, as it may trigger further
  // simplifications.
  template <typename Operands> void addToWorklist(Operands &&operands) {
    for (Value *operand : operands) {
      // If the use count of this operand is now < 2, we re-add the defining
      // operation to the worklist.
      // TODO(riverriddle) This is based on the fact that zero use operations
      // may be deleted, and that single use values often have more
      // canonicalization opportunities.
      if (!operand->use_empty() &&
          std::next(operand->use_begin()) != operand->use_end())
        continue;
      if (auto *defInst = operand->getDefiningOp())
        addToWorklist(defInst);
    }
  }
  /// The low-level pattern matcher.
  RewritePatternMatcher matcher;
  /// This builder is used to create new operations.
  FuncBuilder builder;
  /// The worklist for this transformation keeps track of the operations that
  /// need to be revisited, plus their index in the worklist.  This allows us to
  /// efficiently remove operations from the worklist when they are erased from
  /// the function, even if they aren't the root of a pattern.
  std::vector<Operation *> worklist;
  DenseMap<Operation *, unsigned> worklistMap;
  /// As part of canonicalization, we move constants to the top of the entry
  /// block of the current function and de-duplicate them.  This keeps track of
  /// constants we have done this for.
  DenseMap<std::pair<Attribute, Type>, Operation *> uniquedConstants;
};
}; // end anonymous namespace
/// Perform the rewrites.
void GreedyPatternRewriteDriver::simplifyFunction() {
  // These are scratch vectors used in the constant folding loop below.
  SmallVector<Attribute, 8> operandConstants, resultConstants;
  SmallVector<Value *, 8> originalOperands, resultValues;
  while (!worklist.empty()) {
    auto *op = popFromWorklist();
    // Nulls get added to the worklist when operations are removed, ignore them.
    if (op == nullptr)
      continue;
    // If we have a constant op, unique it into the entry block.
    if (auto constant = op->dyn_cast<ConstantOp>()) {
      // If this constant is dead, remove it, being careful to keep
      // uniquedConstants up to date.
      if (constant.use_empty()) {
        auto it =
            uniquedConstants.find({constant.getValue(), constant.getType()});
        if (it != uniquedConstants.end() && it->second == op)
          uniquedConstants.erase(it);
        constant.erase();
        continue;
      }
      // Check to see if we already have a constant with this type and value:
      auto &entry = uniquedConstants[std::make_pair(constant.getValue(),
                                                    constant.getType())];
      if (entry) {
        // If this constant is already our uniqued one, then leave it alone.
        if (entry == op)
          continue;
        // Otherwise replace this redundant constant with the uniqued one.  We
        // know this is safe because we move constants to the top of the
        // function when they are uniqued, so we know they dominate all uses.
        constant.replaceAllUsesWith(entry->getResult(0));
        constant.erase();
        continue;
      }
      // If we have no entry, then we should unique this constant as the
      // canonical version.  To ensure safe dominance, move the operation to the
      // top of the function.
      entry = op;
      auto &entryBB = builder.getInsertionBlock()->getFunction()->front();
      op->moveBefore(&entryBB, entryBB.begin());
      continue;
    }
    // If the operation has no side effects, and no users, then it is trivially
    // dead - remove it.
    if (op->hasNoSideEffect() && op->use_empty()) {
      op->erase();
      continue;
    }
    // Check to see if any operands to the operation is constant and whether
    // the operation knows how to constant fold itself.
    operandConstants.assign(op->getNumOperands(), Attribute());
    for (unsigned i = 0, e = op->getNumOperands(); i != e; ++i)
      matchPattern(op->getOperand(i), m_Constant(&operandConstants[i]));
    // If this is a commutative binary operation with a constant on the left
    // side move it to the right side.
    if (operandConstants.size() == 2 && operandConstants[0] &&
        !operandConstants[1] && op->isCommutative()) {
      std::swap(op->getInstOperand(0), op->getInstOperand(1));
      std::swap(operandConstants[0], operandConstants[1]);
    }
    // If constant folding was successful, create the result constants, RAUW the
    // operation and remove it.
    resultConstants.clear();
    if (succeeded(op->constantFold(operandConstants, resultConstants))) {
      builder.setInsertionPoint(op);
      // Add the operands to the worklist for visitation.
      addToWorklist(op->getOperands());
      for (unsigned i = 0, e = op->getNumResults(); i != e; ++i) {
        auto *res = op->getResult(i);
        if (res->use_empty()) // ignore dead uses.
          continue;
        // If we already have a canonicalized version of this constant, just
        // reuse it.  Otherwise create a new one.
        Value *cstValue;
        auto it = uniquedConstants.find({resultConstants[i], res->getType()});
        if (it != uniquedConstants.end())
          cstValue = it->second->getResult(0);
        else
          cstValue = create<ConstantOp>(op->getLoc(), res->getType(),
                                        resultConstants[i]);
        // Add all the users of the result to the worklist so we make sure to
        // revisit them.
        //
        // TODO: Add a result->getUsers() iterator.
        for (auto &operand : op->getResult(i)->getUses())
          addToWorklist(operand.getOwner());
        res->replaceAllUsesWith(cstValue);
      }
      assert(op->hasNoSideEffect() && "Constant folded op with side effects?");
      op->erase();
      continue;
    }
    // Otherwise see if we can use the generic folder API to simplify the
    // operation.
    originalOperands.assign(op->operand_begin(), op->operand_end());
    resultValues.clear();
    if (succeeded(op->fold(resultValues))) {
      // If the result was an in-place simplification (e.g. max(x,x,y) ->
      // max(x,y)) then add the original operands to the worklist so we can make
      // sure to revisit them.
      if (resultValues.empty()) {
        // Add the operands back to the worklist as there may be more
        // canonicalization opportunities now.
        addToWorklist(originalOperands);
      } else {
        // Otherwise, the operation is simplified away completely.
        assert(resultValues.size() == op->getNumResults());
        // Notify that we are replacing this operation.
        notifyRootReplaced(op);
        // Replace the result values and erase the operation.
        for (unsigned i = 0, e = resultValues.size(); i != e; ++i) {
          auto *res = op->getResult(i);
          if (!res->use_empty())
            res->replaceAllUsesWith(resultValues[i]);
        }
        notifyOperationRemoved(op);
        op->erase();
      }
      continue;
    }
    // Make sure that any new operations are inserted at this point.
    builder.setInsertionPoint(op);
    // Try to match one of the canonicalization patterns. The rewriter is
    // automatically notified of any necessary changes, so there is nothing else
    // to do here.
    matcher.matchAndRewrite(op);
  }
  uniquedConstants.clear();
}
/// Rewrite the specified function by repeatedly applying the highest benefit
/// patterns in a greedy work-list driven manner.
///
void mlir::applyPatternsGreedily(Function &fn,
                                 OwningRewritePatternList &&patterns) {
  GreedyPatternRewriteDriver driver(fn, std::move(patterns));
  driver.simplifyFunction();
}
  |