-
Notifications
You must be signed in to change notification settings - Fork 10.4k
/
Copy pathARCSequenceOpts.cpp
310 lines (256 loc) · 11.4 KB
/
ARCSequenceOpts.cpp
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
307
308
309
310
//===--- ARCSequenceOpts.cpp ----------------------------------------------===//
//
// This source file is part of the Swift.org open source project
//
// Copyright (c) 2014 - 2017 Apple Inc. and the Swift project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "arc-sequence-opts"
#include "ARCSequenceOpts.h"
#include "swift/SIL/SILBuilder.h"
#include "swift/SIL/SILVisitor.h"
#include "swift/SILOptimizer/Analysis/ARCAnalysis.h"
#include "swift/SILOptimizer/Analysis/AliasAnalysis.h"
#include "swift/SILOptimizer/Analysis/EpilogueARCAnalysis.h"
#include "swift/SILOptimizer/Analysis/LoopAnalysis.h"
#include "swift/SILOptimizer/Analysis/LoopRegionAnalysis.h"
#include "swift/SILOptimizer/Analysis/PostOrderAnalysis.h"
#include "swift/SILOptimizer/Analysis/ProgramTerminationAnalysis.h"
#include "swift/SILOptimizer/Analysis/RCIdentityAnalysis.h"
#include "swift/SILOptimizer/PassManager/Passes.h"
#include "swift/SILOptimizer/PassManager/Transforms.h"
#include "swift/SILOptimizer/Utils/InstOptUtils.h"
#include "swift/SILOptimizer/Utils/LoopUtils.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/PointerUnion.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
using namespace swift;
STATISTIC(NumRefCountOpsRemoved, "Total number of increments removed");
llvm::cl::opt<bool> EnableLoopARC("enable-loop-arc", llvm::cl::init(true));
//===----------------------------------------------------------------------===//
// Code Motion
//===----------------------------------------------------------------------===//
// This routine takes in the ARCMatchingSet \p MatchSet and adds the increments
// and decrements to the delete list.
void ARCPairingContext::optimizeMatchingSet(
ARCMatchingSet &MatchSet,
llvm::SmallVectorImpl<SILInstruction *> &DeadInsts) {
LLVM_DEBUG(llvm::dbgs() << "**** Optimizing Matching Set ****\n");
// Add the old increments to the delete list.
for (SILInstruction *Increment : MatchSet.Increments) {
MadeChange = true;
LLVM_DEBUG(llvm::dbgs() << " Deleting increment: " << *Increment);
DeadInsts.push_back(Increment);
++NumRefCountOpsRemoved;
}
// Add the old decrements to the delete list.
for (SILInstruction *Decrement : MatchSet.Decrements) {
MadeChange = true;
LLVM_DEBUG(llvm::dbgs() << " Deleting decrement: " << *Decrement);
DeadInsts.push_back(Decrement);
++NumRefCountOpsRemoved;
}
}
bool ARCPairingContext::performMatching(
llvm::SmallVectorImpl<SILInstruction *> &DeadInsts) {
bool MatchedPair = false;
LLVM_DEBUG(llvm::dbgs() << "**** Computing ARC Matching Sets for "
<< F.getName() << " ****\n");
/// For each increment that we matched to a decrement, try to match it to a
/// decrement -> increment pair.
for (auto Pair : IncToDecStateMap) {
if (!Pair.has_value())
continue;
SILInstruction *Increment = Pair->first;
if (!Increment)
continue; // blotted
LLVM_DEBUG(llvm::dbgs() << "Constructing Matching Set For: " << *Increment);
ARCMatchingSetBuilder Builder(DecToIncStateMap, IncToDecStateMap, RCIA);
Builder.init(Increment);
if (Builder.matchUpIncDecSetsForPtr()) {
MatchedPair |= Builder.matchedPair();
auto &Set = Builder.getResult();
for (auto *I : Set.Increments)
IncToDecStateMap.erase(I);
for (auto *I : Set.Decrements)
DecToIncStateMap.erase(I);
optimizeMatchingSet(Set, DeadInsts);
}
}
return MatchedPair;
}
//===----------------------------------------------------------------------===//
// Loop ARC
//===----------------------------------------------------------------------===//
void LoopARCPairingContext::runOnLoop(SILLoop *L) {
auto *Region = LRFI->getRegion(L);
if (processRegion(Region, false, false)) {
// We do not recompute for now since we only look at the top function level
// for post dominating releases.
processRegion(Region, true, false);
}
// Now that we have finished processing the loop, summarize the loop.
Evaluator.summarizeLoop(Region);
}
void LoopARCPairingContext::runOnFunction(SILFunction *F) {
if (processRegion(LRFI->getTopLevelRegion(), false, false)) {
// We recompute the final post dom release since we may have moved the final
// post dominated releases.
processRegion(LRFI->getTopLevelRegion(), true, true);
}
}
bool LoopARCPairingContext::processRegion(const LoopRegion *Region,
bool FreezePostDomReleases,
bool RecomputePostDomReleases) {
llvm::SmallVector<SILInstruction *, 8> DeadInsts;
// We have already summarized all subloops of this loop. Now summarize our
// blocks so that we only visit interesting instructions.
Evaluator.summarizeSubregionBlocks(Region);
bool MadeChange = false;
bool NestingDetected = false;
bool MatchedPair = false;
do {
NestingDetected = Evaluator.runOnLoop(Region, FreezePostDomReleases,
RecomputePostDomReleases);
MatchedPair = Context.performMatching(DeadInsts);
if (!DeadInsts.empty()) {
LLVM_DEBUG(llvm::dbgs() << "Removing dead interesting insts!\n");
do {
SILInstruction *I = DeadInsts.pop_back_val();
LLVM_DEBUG(llvm::dbgs() << " " << *I);
Evaluator.removeInterestingInst(I);
I->eraseFromParent();
} while (!DeadInsts.empty());
}
MadeChange |= MatchedPair;
Evaluator.saveMatchingInfo(Region);
Evaluator.clearLoopState(Region);
Context.DecToIncStateMap.clear();
Context.IncToDecStateMap.clear();
Evaluator.clearSetFactory();
// This ensures we only ever recompute post dominating releases on the first
// iteration.
RecomputePostDomReleases = false;
} while (NestingDetected && MatchedPair);
return MadeChange;
}
//===----------------------------------------------------------------------===//
// Non Loop Optimizer
//===----------------------------------------------------------------------===//
static bool
processFunctionWithoutLoopSupport(SILFunction &F, bool FreezePostDomReleases,
AliasAnalysis *AA, PostOrderAnalysis *POTA,
RCIdentityFunctionInfo *RCIA,
EpilogueARCFunctionInfo *EAFI,
ProgramTerminationFunctionInfo *PTFI) {
// GlobalARCOpts seems to be taking up a lot of compile time when running on
// globalinit_func. Since that is not *that* interesting from an ARC
// perspective (i.e. no ref count operations in a loop), disable it on such
// functions temporarily in order to unblock others. This should be removed.
if (F.isGlobalInitOnceFunction())
return false;
LLVM_DEBUG(llvm::dbgs() << "***** Processing " << F.getName() << " *****\n");
bool Changed = false;
BlockARCPairingContext Context(F, AA, POTA, RCIA, EAFI, PTFI);
// Until we do not remove any instructions or have nested increments,
// decrements...
while (true) {
// Compute matching sets of increments, decrements, and their insertion
// points.
//
// We need to blot pointers we remove after processing an individual pointer
// so we don't process pairs after we have paired them up. Thus we pass in a
// lambda that performs the work for us.
bool ShouldRunAgain = Context.run(FreezePostDomReleases);
Changed |= Context.madeChange();
// If we did not remove any instructions or have any nested increments, do
// not perform another iteration.
if (!ShouldRunAgain)
break;
// Otherwise, perform another iteration.
LLVM_DEBUG(llvm::dbgs() << "\n<<< Made a Change! "
"Reprocessing Function! >>>\n");
}
LLVM_DEBUG(llvm::dbgs() << "\n");
// Return true if we moved or deleted any instructions.
return Changed;
}
//===----------------------------------------------------------------------===//
// Loop Optimizer
//===----------------------------------------------------------------------===//
static bool processFunctionWithLoopSupport(
SILFunction &F, AliasAnalysis *AA, PostOrderAnalysis *POTA,
LoopRegionFunctionInfo *LRFI, SILLoopInfo *LI, RCIdentityFunctionInfo *RCFI,
EpilogueARCFunctionInfo *EAFI, ProgramTerminationFunctionInfo *PTFI) {
// GlobalARCOpts seems to be taking up a lot of compile time when running on
// globalinit_func. Since that is not *that* interesting from an ARC
// perspective (i.e. no ref count operations in a loop), disable it on such
// functions temporarily in order to unblock others. This should be removed.
if (F.isGlobalInitOnceFunction())
return false;
LLVM_DEBUG(llvm::dbgs() << "***** Processing " << F.getName() << " *****\n");
LoopARCPairingContext Context(F, AA, LRFI, LI, RCFI, EAFI, PTFI);
return Context.process();
}
//===----------------------------------------------------------------------===//
// Top Level Driver
//===----------------------------------------------------------------------===//
namespace {
class ARCSequenceOpts : public SILFunctionTransform {
/// The entry point to the transformation.
void run() override {
auto *F = getFunction();
// If ARC optimizations are disabled, don't optimize anything and bail.
if (!getOptions().EnableARCOptimizations)
return;
// FIXME: We should support ownership.
if (F->hasOwnership())
return;
if (!EnableLoopARC) {
auto *AA = getAnalysis<AliasAnalysis>(F);
auto *POTA = getAnalysis<PostOrderAnalysis>();
auto *RCFI = getAnalysis<RCIdentityAnalysis>()->get(F);
auto *EAFI = getAnalysis<EpilogueARCAnalysis>()->get(F);
ProgramTerminationFunctionInfo PTFI(F);
if (processFunctionWithoutLoopSupport(*F, false, AA, POTA, RCFI, EAFI, &PTFI)) {
processFunctionWithoutLoopSupport(*F, true, AA, POTA, RCFI, EAFI, &PTFI);
invalidateAnalysis(SILAnalysis::InvalidationKind::CallsAndInstructions);
}
return;
}
auto *LA = getAnalysis<SILLoopAnalysis>();
auto *LI = LA->get(F);
auto *DA = getAnalysis<DominanceAnalysis>();
auto *DI = DA->get(F);
// Canonicalize the loops, invalidating if we need to.
if (canonicalizeAllLoops(DI, LI)) {
// We preserve loop info and the dominator tree.
DA->lockInvalidation();
LA->lockInvalidation();
PM->invalidateAnalysis(F, SILAnalysis::InvalidationKind::FunctionBody);
DA->unlockInvalidation();
LA->unlockInvalidation();
}
auto *AA = getAnalysis<AliasAnalysis>(F);
auto *POTA = getAnalysis<PostOrderAnalysis>();
auto *RCFI = getAnalysis<RCIdentityAnalysis>()->get(F);
auto *EAFI = getAnalysis<EpilogueARCAnalysis>()->get(F);
auto *LRFI = getAnalysis<LoopRegionAnalysis>()->get(F);
ProgramTerminationFunctionInfo PTFI(F);
if (processFunctionWithLoopSupport(*F, AA, POTA, LRFI, LI, RCFI, EAFI, &PTFI)) {
invalidateAnalysis(SILAnalysis::InvalidationKind::CallsAndInstructions);
}
}
};
} // end anonymous namespace
SILTransform *swift::createARCSequenceOpts() {
return new ARCSequenceOpts();
}