1 //===- LoopVectorize.h ------------------------------------------*- C++ -*-===//
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 is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
10 // and generates target-independent LLVM-IR.
11 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
12 // of instructions in order to estimate the profitability of vectorization.
13 //
14 // The loop vectorizer combines consecutive loop iterations into a single
15 // 'wide' iteration. After this transformation the index is incremented
16 // by the SIMD vector width, and not by one.
17 //
18 // This pass has three parts:
19 // 1. The main loop pass that drives the different parts.
20 // 2. LoopVectorizationLegality - A unit that checks for the legality
21 //    of the vectorization.
22 // 3. InnerLoopVectorizer - A unit that performs the actual
23 //    widening of instructions.
24 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
25 //    of vectorization. It decides on the optimal vector width, which
26 //    can be one, if vectorization is not profitable.
27 //
28 // There is a development effort going on to migrate loop vectorizer to the
29 // VPlan infrastructure and to introduce outer loop vectorization support (see
30 // docs/Proposal/VectorizationPlan.rst and
31 // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this
32 // purpose, we temporarily introduced the VPlan-native vectorization path: an
33 // alternative vectorization path that is natively implemented on top of the
34 // VPlan infrastructure. See EnableVPlanNativePath for enabling.
35 //
36 //===----------------------------------------------------------------------===//
37 //
38 // The reduction-variable vectorization is based on the paper:
39 //  D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
40 //
41 // Variable uniformity checks are inspired by:
42 //  Karrenberg, R. and Hack, S. Whole Function Vectorization.
43 //
44 // The interleaved access vectorization is based on the paper:
45 //  Dorit Nuzman, Ira Rosen and Ayal Zaks.  Auto-Vectorization of Interleaved
46 //  Data for SIMD
47 //
48 // Other ideas/concepts are from:
49 //  A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
50 //
51 //  S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua.  An Evaluation of
52 //  Vectorizing Compilers.
53 //
54 //===----------------------------------------------------------------------===//
55 
56 #ifndef LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZE_H
57 #define LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZE_H
58 
59 #include "llvm/Analysis/AliasAnalysis.h"
60 #include "llvm/IR/PassManager.h"
61 #include "llvm/Support/CommandLine.h"
62 #include <functional>
63 
64 namespace llvm {
65 
66 class AssumptionCache;
67 class BlockFrequencyInfo;
68 class DemandedBits;
69 class DominatorTree;
70 class Function;
71 class Loop;
72 class LoopAccessInfo;
73 class LoopInfo;
74 class OptimizationRemarkEmitter;
75 class ProfileSummaryInfo;
76 class ScalarEvolution;
77 class TargetLibraryInfo;
78 class TargetTransformInfo;
79 
80 extern cl::opt<bool> EnableLoopInterleaving;
81 extern cl::opt<bool> EnableLoopVectorization;
82 
83 struct LoopVectorizeOptions {
84   /// If false, consider all loops for interleaving.
85   /// If true, only loops that explicitly request interleaving are considered.
86   bool InterleaveOnlyWhenForced;
87 
88   /// If false, consider all loops for vectorization.
89   /// If true, only loops that explicitly request vectorization are considered.
90   bool VectorizeOnlyWhenForced;
91 
92   /// The current defaults when creating the pass with no arguments are:
93   /// EnableLoopInterleaving = true and EnableLoopVectorization = true. This
94   /// means that interleaving default is consistent with the cl::opt flag, while
95   /// vectorization is not.
96   /// FIXME: The default for EnableLoopVectorization in the cl::opt should be
97   /// set to true, and the corresponding change to account for this be made in
98   /// opt.cpp. The initializations below will become:
99   /// InterleaveOnlyWhenForced(!EnableLoopInterleaving)
100   /// VectorizeOnlyWhenForced(!EnableLoopVectorization).
101   LoopVectorizeOptions()
102       : InterleaveOnlyWhenForced(false), VectorizeOnlyWhenForced(false) {}
103   LoopVectorizeOptions(bool InterleaveOnlyWhenForced,
104                        bool VectorizeOnlyWhenForced)
105       : InterleaveOnlyWhenForced(InterleaveOnlyWhenForced),
106         VectorizeOnlyWhenForced(VectorizeOnlyWhenForced) {}
107 
108   LoopVectorizeOptions &setInterleaveOnlyWhenForced(bool Value) {
109     InterleaveOnlyWhenForced = Value;
110     return *this;
111   }
112 
113   LoopVectorizeOptions &setVectorizeOnlyWhenForced(bool Value) {
114     VectorizeOnlyWhenForced = Value;
115     return *this;
116   }
117 };
118 
119 /// The LoopVectorize Pass.
120 struct LoopVectorizePass : public PassInfoMixin<LoopVectorizePass> {
121   /// If false, consider all loops for interleaving.
122   /// If true, only loops that explicitly request interleaving are considered.
123   bool InterleaveOnlyWhenForced;
124 
125   /// If false, consider all loops for vectorization.
126   /// If true, only loops that explicitly request vectorization are considered.
127   bool VectorizeOnlyWhenForced;
128 
129   LoopVectorizePass(LoopVectorizeOptions Opts = {})
130       : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced),
131         VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced) {}
132 
133   ScalarEvolution *SE;
134   LoopInfo *LI;
135   TargetTransformInfo *TTI;
136   DominatorTree *DT;
137   BlockFrequencyInfo *BFI;
138   TargetLibraryInfo *TLI;
139   DemandedBits *DB;
140   AliasAnalysis *AA;
141   AssumptionCache *AC;
142   std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
143   OptimizationRemarkEmitter *ORE;
144   ProfileSummaryInfo *PSI;
145 
146   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
147 
148   // Shim for old PM.
149   bool runImpl(Function &F, ScalarEvolution &SE_, LoopInfo &LI_,
150                TargetTransformInfo &TTI_, DominatorTree &DT_,
151                BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_,
152                DemandedBits &DB_, AliasAnalysis &AA_, AssumptionCache &AC_,
153                std::function<const LoopAccessInfo &(Loop &)> &GetLAA_,
154                OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_);
155 
156   bool processLoop(Loop *L);
157 };
158 
159 /// Reports a vectorization failure: print \p DebugMsg for debugging
160 /// purposes along with the corresponding optimization remark \p RemarkName.
161 /// If \p I is passed, it is an instruction that prevents vectorization.
162 /// Otherwise, the loop \p TheLoop is used for the location of the remark.
163 void reportVectorizationFailure(const StringRef DebugMsg,
164     const StringRef OREMsg, const StringRef ORETag,
165     OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I = nullptr);
166 
167 } // end namespace llvm
168 
169 #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZE_H
170