blob: 5f332e3cac1623e9ef60653c4eac712d958d5113 [file] [log] [blame]
Andrew Scull5e1ddfa2018-08-14 10:06:54 +01001//===- llvm/Analysis/LoopUnrollAnalyzer.h - Loop Unroll Analyzer-*- C++ -*-===//
2//
Andrew Walbran16937d02019-10-22 13:54:20 +01003// 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
Andrew Scull5e1ddfa2018-08-14 10:06:54 +01006//
7//===----------------------------------------------------------------------===//
8//
9// This file implements UnrolledInstAnalyzer class. It's used for predicting
10// potential effects that loop unrolling might have, such as enabling constant
11// propagation and other optimizations.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_ANALYSIS_LOOPUNROLLANALYZER_H
16#define LLVM_ANALYSIS_LOOPUNROLLANALYZER_H
17
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/ScalarEvolutionExpressions.h"
20#include "llvm/IR/InstVisitor.h"
21
22// This class is used to get an estimate of the optimization effects that we
23// could get from complete loop unrolling. It comes from the fact that some
24// loads might be replaced with concrete constant values and that could trigger
25// a chain of instruction simplifications.
26//
27// E.g. we might have:
28// int a[] = {0, 1, 0};
29// v = 0;
30// for (i = 0; i < 3; i ++)
31// v += b[i]*a[i];
32// If we completely unroll the loop, we would get:
33// v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
34// Which then will be simplified to:
35// v = b[0]* 0 + b[1]* 1 + b[2]* 0
36// And finally:
37// v = b[1]
38namespace llvm {
39class UnrolledInstAnalyzer : private InstVisitor<UnrolledInstAnalyzer, bool> {
40 typedef InstVisitor<UnrolledInstAnalyzer, bool> Base;
41 friend class InstVisitor<UnrolledInstAnalyzer, bool>;
42 struct SimplifiedAddress {
43 Value *Base = nullptr;
44 ConstantInt *Offset = nullptr;
45 };
46
47public:
48 UnrolledInstAnalyzer(unsigned Iteration,
49 DenseMap<Value *, Constant *> &SimplifiedValues,
50 ScalarEvolution &SE, const Loop *L)
51 : SimplifiedValues(SimplifiedValues), SE(SE), L(L) {
52 IterationNumber = SE.getConstant(APInt(64, Iteration));
53 }
54
55 // Allow access to the initial visit method.
56 using Base::visit;
57
58private:
Andrew Scullcdfcccc2018-10-05 20:58:37 +010059 /// A cache of pointer bases and constant-folded offsets corresponding
Andrew Scull5e1ddfa2018-08-14 10:06:54 +010060 /// to GEP (or derived from GEP) instructions.
61 ///
62 /// In order to find the base pointer one needs to perform non-trivial
63 /// traversal of the corresponding SCEV expression, so it's good to have the
64 /// results saved.
65 DenseMap<Value *, SimplifiedAddress> SimplifiedAddresses;
66
Andrew Scullcdfcccc2018-10-05 20:58:37 +010067 /// SCEV expression corresponding to number of currently simulated
Andrew Scull5e1ddfa2018-08-14 10:06:54 +010068 /// iteration.
69 const SCEV *IterationNumber;
70
Andrew Scullcdfcccc2018-10-05 20:58:37 +010071 /// A Value->Constant map for keeping values that we managed to
Andrew Scull5e1ddfa2018-08-14 10:06:54 +010072 /// constant-fold on the given iteration.
73 ///
74 /// While we walk the loop instructions, we build up and maintain a mapping
75 /// of simplified values specific to this iteration. The idea is to propagate
76 /// any special information we have about loads that can be replaced with
77 /// constants after complete unrolling, and account for likely simplifications
78 /// post-unrolling.
79 DenseMap<Value *, Constant *> &SimplifiedValues;
80
81 ScalarEvolution &SE;
82 const Loop *L;
83
84 bool simplifyInstWithSCEV(Instruction *I);
85
86 bool visitInstruction(Instruction &I) { return simplifyInstWithSCEV(&I); }
87 bool visitBinaryOperator(BinaryOperator &I);
88 bool visitLoad(LoadInst &I);
89 bool visitCastInst(CastInst &I);
90 bool visitCmpInst(CmpInst &I);
91 bool visitPHINode(PHINode &PN);
92};
93}
94#endif