1 | //===- LoopInstSimplify.cpp - Loop Instruction Simplification Pass --------===// |
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 pass performs lightweight instruction simplification on loop bodies. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Transforms/Scalar/LoopInstSimplify.h" |
14 | #include "llvm/ADT/STLExtras.h" |
15 | #include "llvm/ADT/SmallPtrSet.h" |
16 | #include "llvm/ADT/SmallVector.h" |
17 | #include "llvm/ADT/Statistic.h" |
18 | #include "llvm/Analysis/AssumptionCache.h" |
19 | #include "llvm/Analysis/InstructionSimplify.h" |
20 | #include "llvm/Analysis/LoopInfo.h" |
21 | #include "llvm/Analysis/LoopIterator.h" |
22 | #include "llvm/Analysis/LoopPass.h" |
23 | #include "llvm/Analysis/MemorySSA.h" |
24 | #include "llvm/Analysis/MemorySSAUpdater.h" |
25 | #include "llvm/Analysis/TargetLibraryInfo.h" |
26 | #include "llvm/IR/BasicBlock.h" |
27 | #include "llvm/IR/Dominators.h" |
28 | #include "llvm/IR/Instruction.h" |
29 | #include "llvm/IR/Instructions.h" |
30 | #include "llvm/Support/Casting.h" |
31 | #include "llvm/Transforms/Utils/Local.h" |
32 | #include "llvm/Transforms/Utils/LoopUtils.h" |
33 | #include <optional> |
34 | #include <utility> |
35 | |
36 | using namespace llvm; |
37 | |
38 | #define DEBUG_TYPE "loop-instsimplify" |
39 | |
40 | STATISTIC(NumSimplified, "Number of redundant instructions simplified" ); |
41 | |
42 | static bool simplifyLoopInst(Loop &L, DominatorTree &DT, LoopInfo &LI, |
43 | AssumptionCache &AC, const TargetLibraryInfo &TLI, |
44 | MemorySSAUpdater *MSSAU) { |
45 | const DataLayout &DL = L.getHeader()->getDataLayout(); |
46 | SimplifyQuery SQ(DL, &TLI, &DT, &AC); |
47 | |
48 | // On the first pass over the loop body we try to simplify every instruction. |
49 | // On subsequent passes, we can restrict this to only simplifying instructions |
50 | // where the inputs have been updated. We end up needing two sets: one |
51 | // containing the instructions we are simplifying in *this* pass, and one for |
52 | // the instructions we will want to simplify in the *next* pass. We use |
53 | // pointers so we can swap between two stably allocated sets. |
54 | SmallPtrSet<const Instruction *, 8> S1, S2, *ToSimplify = &S1, *Next = &S2; |
55 | |
56 | // Track the PHI nodes that have already been visited during each iteration so |
57 | // that we can identify when it is necessary to iterate. |
58 | SmallPtrSet<PHINode *, 4> VisitedPHIs; |
59 | |
60 | // While simplifying we may discover dead code or cause code to become dead. |
61 | // Keep track of all such instructions and we will delete them at the end. |
62 | SmallVector<WeakTrackingVH, 8> DeadInsts; |
63 | |
64 | // First we want to create an RPO traversal of the loop body. By processing in |
65 | // RPO we can ensure that definitions are processed prior to uses (for non PHI |
66 | // uses) in all cases. This ensures we maximize the simplifications in each |
67 | // iteration over the loop and minimizes the possible causes for continuing to |
68 | // iterate. |
69 | LoopBlocksRPO RPOT(&L); |
70 | RPOT.perform(LI: &LI); |
71 | MemorySSA *MSSA = MSSAU ? MSSAU->getMemorySSA() : nullptr; |
72 | |
73 | bool Changed = false; |
74 | for (;;) { |
75 | if (MSSAU && VerifyMemorySSA) |
76 | MSSA->verifyMemorySSA(); |
77 | for (BasicBlock *BB : RPOT) { |
78 | for (Instruction &I : *BB) { |
79 | if (auto *PI = dyn_cast<PHINode>(Val: &I)) |
80 | VisitedPHIs.insert(Ptr: PI); |
81 | |
82 | if (I.use_empty()) { |
83 | if (isInstructionTriviallyDead(I: &I, TLI: &TLI)) |
84 | DeadInsts.push_back(Elt: &I); |
85 | continue; |
86 | } |
87 | |
88 | // We special case the first iteration which we can detect due to the |
89 | // empty `ToSimplify` set. |
90 | bool IsFirstIteration = ToSimplify->empty(); |
91 | |
92 | if (!IsFirstIteration && !ToSimplify->count(Ptr: &I)) |
93 | continue; |
94 | |
95 | Value *V = simplifyInstruction(I: &I, Q: SQ.getWithInstruction(I: &I)); |
96 | if (!V || !LI.replacementPreservesLCSSAForm(From: &I, To: V)) |
97 | continue; |
98 | |
99 | for (Use &U : llvm::make_early_inc_range(Range: I.uses())) { |
100 | auto *UserI = cast<Instruction>(Val: U.getUser()); |
101 | U.set(V); |
102 | |
103 | // Do not bother dealing with unreachable code. |
104 | if (!DT.isReachableFromEntry(A: UserI->getParent())) |
105 | continue; |
106 | |
107 | // If the instruction is used by a PHI node we have already processed |
108 | // we'll need to iterate on the loop body to converge, so add it to |
109 | // the next set. |
110 | if (auto *UserPI = dyn_cast<PHINode>(Val: UserI)) |
111 | if (VisitedPHIs.count(Ptr: UserPI)) { |
112 | Next->insert(Ptr: UserPI); |
113 | continue; |
114 | } |
115 | |
116 | // If we are only simplifying targeted instructions and the user is an |
117 | // instruction in the loop body, add it to our set of targeted |
118 | // instructions. Because we process defs before uses (outside of PHIs) |
119 | // we won't have visited it yet. |
120 | // |
121 | // We also skip any uses outside of the loop being simplified. Those |
122 | // should always be PHI nodes due to LCSSA form, and we don't want to |
123 | // try to simplify those away. |
124 | assert((L.contains(UserI) || isa<PHINode>(UserI)) && |
125 | "Uses outside the loop should be PHI nodes due to LCSSA!" ); |
126 | if (!IsFirstIteration && L.contains(Inst: UserI)) |
127 | ToSimplify->insert(Ptr: UserI); |
128 | } |
129 | |
130 | if (MSSAU) |
131 | if (Instruction *SimpleI = dyn_cast_or_null<Instruction>(Val: V)) |
132 | if (MemoryAccess *MA = MSSA->getMemoryAccess(I: &I)) |
133 | if (MemoryAccess *ReplacementMA = MSSA->getMemoryAccess(I: SimpleI)) |
134 | MA->replaceAllUsesWith(V: ReplacementMA); |
135 | |
136 | assert(I.use_empty() && "Should always have replaced all uses!" ); |
137 | if (isInstructionTriviallyDead(I: &I, TLI: &TLI)) |
138 | DeadInsts.push_back(Elt: &I); |
139 | ++NumSimplified; |
140 | Changed = true; |
141 | } |
142 | } |
143 | |
144 | // Delete any dead instructions found thus far now that we've finished an |
145 | // iteration over all instructions in all the loop blocks. |
146 | if (!DeadInsts.empty()) { |
147 | Changed = true; |
148 | RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI: &TLI, MSSAU); |
149 | } |
150 | |
151 | if (MSSAU && VerifyMemorySSA) |
152 | MSSA->verifyMemorySSA(); |
153 | |
154 | // If we never found a PHI that needs to be simplified in the next |
155 | // iteration, we're done. |
156 | if (Next->empty()) |
157 | break; |
158 | |
159 | // Otherwise, put the next set in place for the next iteration and reset it |
160 | // and the visited PHIs for that iteration. |
161 | std::swap(a&: Next, b&: ToSimplify); |
162 | Next->clear(); |
163 | VisitedPHIs.clear(); |
164 | DeadInsts.clear(); |
165 | } |
166 | |
167 | return Changed; |
168 | } |
169 | |
170 | PreservedAnalyses LoopInstSimplifyPass::run(Loop &L, LoopAnalysisManager &AM, |
171 | LoopStandardAnalysisResults &AR, |
172 | LPMUpdater &) { |
173 | std::optional<MemorySSAUpdater> MSSAU; |
174 | if (AR.MSSA) { |
175 | MSSAU = MemorySSAUpdater(AR.MSSA); |
176 | if (VerifyMemorySSA) |
177 | AR.MSSA->verifyMemorySSA(); |
178 | } |
179 | if (!simplifyLoopInst(L, DT&: AR.DT, LI&: AR.LI, AC&: AR.AC, TLI: AR.TLI, |
180 | MSSAU: MSSAU ? &*MSSAU : nullptr)) |
181 | return PreservedAnalyses::all(); |
182 | |
183 | auto PA = getLoopPassPreservedAnalyses(); |
184 | PA.preserveSet<CFGAnalyses>(); |
185 | if (AR.MSSA) |
186 | PA.preserve<MemorySSAAnalysis>(); |
187 | return PA; |
188 | } |
189 | |