[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / Transforms / Scalar / BDCE.cpp
blob1fa2c75b0f42ac186b1b73cd72a55b2d46799f9e
1 //===---- BDCE.cpp - Bit-tracking dead code elimination -------------------===//
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 file implements the Bit-Tracking Dead Code Elimination pass. Some
10 // instructions (shifts, some ands, ors, etc.) kill some of their input bits.
11 // We track these dead bits and remove instructions that compute only these
12 // dead bits. We also simplify sext that generates unused extension bits,
13 // converting it to a zext.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/Scalar/BDCE.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/DemandedBits.h"
22 #include "llvm/Analysis/GlobalsModRef.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/IR/InstIterator.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Transforms/Utils/Local.h"
29 using namespace llvm;
31 #define DEBUG_TYPE "bdce"
33 STATISTIC(NumRemoved, "Number of instructions removed (unused)");
34 STATISTIC(NumSimplified, "Number of instructions trivialized (dead bits)");
35 STATISTIC(NumSExt2ZExt,
36 "Number of sign extension instructions converted to zero extension");
38 /// If an instruction is trivialized (dead), then the chain of users of that
39 /// instruction may need to be cleared of assumptions that can no longer be
40 /// guaranteed correct.
41 static void clearAssumptionsOfUsers(Instruction *I, DemandedBits &DB) {
42 assert(I->getType()->isIntOrIntVectorTy() &&
43 "Trivializing a non-integer value?");
45 // Initialize the worklist with eligible direct users.
46 SmallPtrSet<Instruction *, 16> Visited;
47 SmallVector<Instruction *, 16> WorkList;
48 for (User *JU : I->users()) {
49 // If all bits of a user are demanded, then we know that nothing below that
50 // in the def-use chain needs to be changed.
51 auto *J = dyn_cast<Instruction>(JU);
52 if (J && J->getType()->isIntOrIntVectorTy() &&
53 !DB.getDemandedBits(J).isAllOnes()) {
54 Visited.insert(J);
55 WorkList.push_back(J);
58 // Note that we need to check for non-int types above before asking for
59 // demanded bits. Normally, the only way to reach an instruction with an
60 // non-int type is via an instruction that has side effects (or otherwise
61 // will demand its input bits). However, if we have a readnone function
62 // that returns an unsized type (e.g., void), we must avoid asking for the
63 // demanded bits of the function call's return value. A void-returning
64 // readnone function is always dead (and so we can stop walking the use/def
65 // chain here), but the check is necessary to avoid asserting.
68 // DFS through subsequent users while tracking visits to avoid cycles.
69 while (!WorkList.empty()) {
70 Instruction *J = WorkList.pop_back_val();
72 // NSW, NUW, and exact are based on operands that might have changed.
73 J->dropPoisonGeneratingFlags();
75 // We do not have to worry about llvm.assume or range metadata:
76 // 1. llvm.assume demands its operand, so trivializing can't change it.
77 // 2. range metadata only applies to memory accesses which demand all bits.
79 for (User *KU : J->users()) {
80 // If all bits of a user are demanded, then we know that nothing below
81 // that in the def-use chain needs to be changed.
82 auto *K = dyn_cast<Instruction>(KU);
83 if (K && Visited.insert(K).second && K->getType()->isIntOrIntVectorTy() &&
84 !DB.getDemandedBits(K).isAllOnes())
85 WorkList.push_back(K);
90 static bool bitTrackingDCE(Function &F, DemandedBits &DB) {
91 SmallVector<Instruction*, 128> Worklist;
92 bool Changed = false;
93 for (Instruction &I : instructions(F)) {
94 // If the instruction has side effects and no non-dbg uses,
95 // skip it. This way we avoid computing known bits on an instruction
96 // that will not help us.
97 if (I.mayHaveSideEffects() && I.use_empty())
98 continue;
100 // Remove instructions that are dead, either because they were not reached
101 // during analysis or have no demanded bits.
102 if (DB.isInstructionDead(&I) ||
103 (I.getType()->isIntOrIntVectorTy() && DB.getDemandedBits(&I).isZero() &&
104 wouldInstructionBeTriviallyDead(&I))) {
105 Worklist.push_back(&I);
106 Changed = true;
107 continue;
110 // Convert SExt into ZExt if none of the extension bits is required
111 if (SExtInst *SE = dyn_cast<SExtInst>(&I)) {
112 APInt Demanded = DB.getDemandedBits(SE);
113 const uint32_t SrcBitSize = SE->getSrcTy()->getScalarSizeInBits();
114 auto *const DstTy = SE->getDestTy();
115 const uint32_t DestBitSize = DstTy->getScalarSizeInBits();
116 if (Demanded.countl_zero() >= (DestBitSize - SrcBitSize)) {
117 clearAssumptionsOfUsers(SE, DB);
118 IRBuilder<> Builder(SE);
119 I.replaceAllUsesWith(
120 Builder.CreateZExt(SE->getOperand(0), DstTy, SE->getName()));
121 Worklist.push_back(SE);
122 Changed = true;
123 NumSExt2ZExt++;
124 continue;
128 for (Use &U : I.operands()) {
129 // DemandedBits only detects dead integer uses.
130 if (!U->getType()->isIntOrIntVectorTy())
131 continue;
133 if (!isa<Instruction>(U) && !isa<Argument>(U))
134 continue;
136 if (!DB.isUseDead(&U))
137 continue;
139 LLVM_DEBUG(dbgs() << "BDCE: Trivializing: " << U << " (all bits dead)\n");
141 clearAssumptionsOfUsers(&I, DB);
143 // Substitute all uses with zero. In theory we could use `freeze poison`
144 // instead, but that seems unlikely to be profitable.
145 U.set(ConstantInt::get(U->getType(), 0));
146 ++NumSimplified;
147 Changed = true;
151 for (Instruction *&I : llvm::reverse(Worklist)) {
152 salvageDebugInfo(*I);
153 I->dropAllReferences();
156 for (Instruction *&I : Worklist) {
157 ++NumRemoved;
158 I->eraseFromParent();
161 return Changed;
164 PreservedAnalyses BDCEPass::run(Function &F, FunctionAnalysisManager &AM) {
165 auto &DB = AM.getResult<DemandedBitsAnalysis>(F);
166 if (!bitTrackingDCE(F, DB))
167 return PreservedAnalyses::all();
169 PreservedAnalyses PA;
170 PA.preserveSet<CFGAnalyses>();
171 return PA;