1 | //===- DeadCodeElimination.cpp - Eliminate dead iteration ----------------===// |
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 | // The polyhedral dead code elimination pass analyses a SCoP to eliminate |
10 | // statement instances that can be proven dead. |
11 | // As a consequence, the code generated for this SCoP may execute a statement |
12 | // less often. This means, a statement may be executed only in certain loop |
13 | // iterations or it may not even be part of the generated code at all. |
14 | // |
15 | // This code: |
16 | // |
17 | // for (i = 0; i < N; i++) |
18 | // arr[i] = 0; |
19 | // for (i = 0; i < N; i++) |
20 | // arr[i] = 10; |
21 | // for (i = 0; i < N; i++) |
22 | // arr[i] = i; |
23 | // |
24 | // is e.g. simplified to: |
25 | // |
26 | // for (i = 0; i < N; i++) |
27 | // arr[i] = i; |
28 | // |
29 | // The idea and the algorithm used was first implemented by Sven Verdoolaege in |
30 | // the 'ppcg' tool. |
31 | // |
32 | //===----------------------------------------------------------------------===// |
33 | |
34 | #include "polly/DeadCodeElimination.h" |
35 | #include "polly/DependenceInfo.h" |
36 | #include "polly/LinkAllPasses.h" |
37 | #include "polly/Options.h" |
38 | #include "polly/ScopInfo.h" |
39 | #include "llvm/Support/CommandLine.h" |
40 | #include "isl/isl-noexceptions.h" |
41 | |
42 | using namespace llvm; |
43 | using namespace polly; |
44 | |
45 | namespace { |
46 | |
47 | cl::opt<int> DCEPreciseSteps( |
48 | "polly-dce-precise-steps" , |
49 | cl::desc("The number of precise steps between two approximating " |
50 | "iterations. (A value of -1 schedules another approximation stage " |
51 | "before the actual dead code elimination." ), |
52 | cl::init(Val: -1), cl::cat(PollyCategory)); |
53 | |
54 | class DeadCodeElimWrapperPass final : public ScopPass { |
55 | public: |
56 | static char ID; |
57 | explicit DeadCodeElimWrapperPass() : ScopPass(ID) {} |
58 | |
59 | /// Remove dead iterations from the schedule of @p S. |
60 | bool runOnScop(Scop &S) override; |
61 | |
62 | /// Register all analyses and transformation required. |
63 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
64 | }; |
65 | |
66 | char DeadCodeElimWrapperPass::ID = 0; |
67 | |
68 | /// Return the set of live iterations. |
69 | /// |
70 | /// The set of live iterations are all iterations that write to memory and for |
71 | /// which we can not prove that there will be a later write that _must_ |
72 | /// overwrite the same memory location and is consequently the only one that |
73 | /// is visible after the execution of the SCoP. |
74 | /// |
75 | /// To compute the live outs, we compute for the data-locations that are |
76 | /// must-written to the last statement that touches these locations. On top of |
77 | /// this we add all statements that perform may-write accesses. |
78 | /// |
79 | /// We could be more precise by removing may-write accesses for which we know |
80 | /// that they are overwritten by a must-write after. However, at the moment the |
81 | /// only may-writes we introduce access the full (unbounded) array, such that |
82 | /// bounded write accesses can not overwrite all of the data-locations. As |
83 | /// this means may-writes are in the current situation always live, there is |
84 | /// no point in trying to remove them from the live-out set. |
85 | static isl::union_set getLiveOut(Scop &S) { |
86 | isl::union_map Schedule = S.getSchedule(); |
87 | isl::union_map MustWrites = S.getMustWrites(); |
88 | isl::union_map WriteIterations = MustWrites.reverse(); |
89 | isl::union_map WriteTimes = WriteIterations.apply_range(umap2: Schedule); |
90 | |
91 | isl::union_map LastWriteTimes = WriteTimes.lexmax(); |
92 | isl::union_map LastWriteIterations = |
93 | LastWriteTimes.apply_range(umap2: Schedule.reverse()); |
94 | |
95 | isl::union_set Live = LastWriteIterations.range(); |
96 | isl::union_map MayWrites = S.getMayWrites(); |
97 | Live = Live.unite(uset2: MayWrites.domain()); |
98 | return Live.coalesce(); |
99 | } |
100 | |
101 | /// Performs polyhedral dead iteration elimination by: |
102 | /// o Assuming that the last write to each location is live. |
103 | /// o Following each RAW dependency from a live iteration backwards and adding |
104 | /// that iteration to the live set. |
105 | /// |
106 | /// To ensure the set of live iterations does not get too complex we always |
107 | /// combine a certain number of precise steps with one approximating step that |
108 | /// simplifies the life set with an affine hull. |
109 | static bool runDeadCodeElimination(Scop &S, int PreciseSteps, |
110 | const Dependences &D) { |
111 | if (!D.hasValidDependences()) |
112 | return false; |
113 | |
114 | isl::union_set Live = getLiveOut(S); |
115 | isl::union_map Dep = |
116 | D.getDependences(Kinds: Dependences::TYPE_RAW | Dependences::TYPE_RED); |
117 | Dep = Dep.reverse(); |
118 | |
119 | if (PreciseSteps == -1) |
120 | Live = Live.affine_hull(); |
121 | |
122 | isl::union_set OriginalDomain = S.getDomains(); |
123 | int Steps = 0; |
124 | while (true) { |
125 | Steps++; |
126 | |
127 | isl::union_set = Live.apply(umap: Dep); |
128 | |
129 | if (Extra.is_subset(uset2: Live)) |
130 | break; |
131 | |
132 | Live = Live.unite(uset2: Extra); |
133 | |
134 | if (Steps > PreciseSteps) { |
135 | Steps = 0; |
136 | Live = Live.affine_hull(); |
137 | } |
138 | |
139 | Live = Live.intersect(uset2: OriginalDomain); |
140 | } |
141 | |
142 | Live = Live.coalesce(); |
143 | |
144 | return S.restrictDomains(Domain: Live); |
145 | } |
146 | |
147 | bool DeadCodeElimWrapperPass::runOnScop(Scop &S) { |
148 | auto &DI = getAnalysis<DependenceInfo>(); |
149 | const Dependences &Deps = DI.getDependences(Level: Dependences::AL_Statement); |
150 | |
151 | bool Changed = runDeadCodeElimination(S, PreciseSteps: DCEPreciseSteps, D: Deps); |
152 | |
153 | // FIXME: We can probably avoid the recomputation of all dependences by |
154 | // updating them explicitly. |
155 | if (Changed) |
156 | DI.recomputeDependences(Level: Dependences::AL_Statement); |
157 | |
158 | return false; |
159 | } |
160 | |
161 | void DeadCodeElimWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { |
162 | ScopPass::getAnalysisUsage(AU); |
163 | AU.addRequired<DependenceInfo>(); |
164 | } |
165 | |
166 | } // namespace |
167 | |
168 | Pass *polly::createDeadCodeElimWrapperPass() { |
169 | return new DeadCodeElimWrapperPass(); |
170 | } |
171 | |
172 | llvm::PreservedAnalyses DeadCodeElimPass::run(Scop &S, ScopAnalysisManager &SAM, |
173 | ScopStandardAnalysisResults &SAR, |
174 | SPMUpdater &U) { |
175 | DependenceAnalysis::Result &DA = SAM.getResult<DependenceAnalysis>(IR&: S, ExtraArgs&: SAR); |
176 | const Dependences &Deps = DA.getDependences(Level: Dependences::AL_Statement); |
177 | |
178 | bool Changed = runDeadCodeElimination(S, PreciseSteps: DCEPreciseSteps, D: Deps); |
179 | |
180 | // FIXME: We can probably avoid the recomputation of all dependences by |
181 | // updating them explicitly. |
182 | if (Changed) |
183 | DA.recomputeDependences(Level: Dependences::AL_Statement); |
184 | |
185 | if (!Changed) |
186 | return PreservedAnalyses::all(); |
187 | |
188 | PreservedAnalyses PA; |
189 | PA.preserveSet<AllAnalysesOn<Module>>(); |
190 | PA.preserveSet<AllAnalysesOn<Function>>(); |
191 | PA.preserveSet<AllAnalysesOn<Loop>>(); |
192 | return PA; |
193 | } |
194 | |
195 | INITIALIZE_PASS_BEGIN(DeadCodeElimWrapperPass, "polly-dce" , |
196 | "Polly - Remove dead iterations" , false, false) |
197 | INITIALIZE_PASS_DEPENDENCY(DependenceInfo) |
198 | INITIALIZE_PASS_DEPENDENCY(ScopInfoRegionPass) |
199 | INITIALIZE_PASS_END(DeadCodeElimWrapperPass, "polly-dce" , |
200 | "Polly - Remove dead iterations" , false, false) |
201 | |