1//===-- Sink.cpp - Code Sinking -------------------------------------------===//
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 moves instructions into successor blocks, when possible, so that
10// they aren't executed on paths where their results aren't needed.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/Scalar/Sink.h"
15#include "llvm/ADT/Statistic.h"
16#include "llvm/Analysis/AliasAnalysis.h"
17#include "llvm/Analysis/LoopInfo.h"
18#include "llvm/IR/Dominators.h"
19#include "llvm/IR/IntrinsicInst.h"
20#include "llvm/InitializePasses.h"
21#include "llvm/Support/Debug.h"
22#include "llvm/Support/raw_ostream.h"
23#include "llvm/Transforms/Scalar.h"
24using namespace llvm;
25
26#define DEBUG_TYPE "sink"
27
28STATISTIC(NumSunk, "Number of instructions sunk");
29STATISTIC(NumSinkIter, "Number of sinking iterations");
30
31static bool isSafeToMove(Instruction *Inst, AliasAnalysis &AA,
32 SmallPtrSetImpl<Instruction *> &Stores) {
33
34 if (Inst->mayWriteToMemory()) {
35 Stores.insert(Ptr: Inst);
36 return false;
37 }
38
39 // Don't sink static alloca instructions. CodeGen assumes allocas outside the
40 // entry block are dynamically sized stack objects.
41 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: Inst))
42 if (AI->isStaticAlloca())
43 return false;
44
45 if (LoadInst *L = dyn_cast<LoadInst>(Val: Inst)) {
46 MemoryLocation Loc = MemoryLocation::get(LI: L);
47 for (Instruction *S : Stores)
48 if (isModSet(MRI: AA.getModRefInfo(I: S, OptLoc: Loc)))
49 return false;
50 }
51
52 if (Inst->isTerminator() || isa<PHINode>(Val: Inst) || Inst->isEHPad() ||
53 Inst->mayThrow() || !Inst->willReturn())
54 return false;
55
56 if (auto *Call = dyn_cast<CallBase>(Val: Inst)) {
57 // Convergent operations cannot be made control-dependent on additional
58 // values.
59 if (Call->isConvergent())
60 return false;
61
62 for (Instruction *S : Stores)
63 if (isModSet(MRI: AA.getModRefInfo(I: S, Call)))
64 return false;
65 }
66
67 return true;
68}
69
70/// IsAcceptableTarget - Return true if it is possible to sink the instruction
71/// in the specified basic block.
72static bool IsAcceptableTarget(Instruction *Inst, BasicBlock *SuccToSinkTo,
73 DominatorTree &DT, LoopInfo &LI) {
74 assert(Inst && "Instruction to be sunk is null");
75 assert(SuccToSinkTo && "Candidate sink target is null");
76
77 // It's never legal to sink an instruction into an EH-pad block.
78 if (SuccToSinkTo->isEHPad())
79 return false;
80
81 // If the block has multiple predecessors, this would introduce computation
82 // on different code paths. We could split the critical edge, but for now we
83 // just punt.
84 // FIXME: Split critical edges if not backedges.
85 if (SuccToSinkTo->getUniquePredecessor() != Inst->getParent()) {
86 // We cannot sink a load across a critical edge - there may be stores in
87 // other code paths.
88 if (Inst->mayReadFromMemory() &&
89 !Inst->hasMetadata(KindID: LLVMContext::MD_invariant_load))
90 return false;
91
92 // Don't sink instructions into a loop.
93 Loop *succ = LI.getLoopFor(BB: SuccToSinkTo);
94 Loop *cur = LI.getLoopFor(BB: Inst->getParent());
95 if (succ != nullptr && succ != cur)
96 return false;
97 }
98
99 return true;
100}
101
102/// SinkInstruction - Determine whether it is safe to sink the specified machine
103/// instruction out of its current block into a successor.
104static bool SinkInstruction(Instruction *Inst,
105 SmallPtrSetImpl<Instruction *> &Stores,
106 DominatorTree &DT, LoopInfo &LI, AAResults &AA) {
107
108 // Check if it's safe to move the instruction.
109 if (!isSafeToMove(Inst, AA, Stores))
110 return false;
111
112 // FIXME: This should include support for sinking instructions within the
113 // block they are currently in to shorten the live ranges. We often get
114 // instructions sunk into the top of a large block, but it would be better to
115 // also sink them down before their first use in the block. This xform has to
116 // be careful not to *increase* register pressure though, e.g. sinking
117 // "x = y + z" down if it kills y and z would increase the live ranges of y
118 // and z and only shrink the live range of x.
119
120 // SuccToSinkTo - This is the successor to sink this instruction to, once we
121 // decide.
122 BasicBlock *SuccToSinkTo = nullptr;
123
124 // Find the nearest common dominator of all users as the candidate.
125 BasicBlock *BB = Inst->getParent();
126 for (Use &U : Inst->uses()) {
127 Instruction *UseInst = cast<Instruction>(Val: U.getUser());
128 BasicBlock *UseBlock = UseInst->getParent();
129 if (PHINode *PN = dyn_cast<PHINode>(Val: UseInst)) {
130 // PHI nodes use the operand in the predecessor block, not the block with
131 // the PHI.
132 unsigned Num = PHINode::getIncomingValueNumForOperand(i: U.getOperandNo());
133 UseBlock = PN->getIncomingBlock(i: Num);
134 }
135 // Don't worry about dead users.
136 if (!DT.isReachableFromEntry(A: UseBlock))
137 continue;
138
139 if (SuccToSinkTo)
140 SuccToSinkTo = DT.findNearestCommonDominator(A: SuccToSinkTo, B: UseBlock);
141 else
142 SuccToSinkTo = UseBlock;
143 }
144
145 if (SuccToSinkTo) {
146 // The nearest common dominator may be in a parent loop of BB, which may not
147 // be beneficial. Find an ancestor.
148 while (SuccToSinkTo != BB &&
149 !IsAcceptableTarget(Inst, SuccToSinkTo, DT, LI))
150 SuccToSinkTo = DT.getNode(BB: SuccToSinkTo)->getIDom()->getBlock();
151 if (SuccToSinkTo == BB)
152 SuccToSinkTo = nullptr;
153 }
154
155 // If we couldn't find a block to sink to, ignore this instruction.
156 if (!SuccToSinkTo)
157 return false;
158
159 LLVM_DEBUG(dbgs() << "Sink" << *Inst << " (";
160 Inst->getParent()->printAsOperand(dbgs(), false); dbgs() << " -> ";
161 SuccToSinkTo->printAsOperand(dbgs(), false); dbgs() << ")\n");
162
163 // The current location of Inst dominates all uses, thus it must dominate
164 // SuccToSinkTo, which is on the IDom chain between the nearest common
165 // dominator to all uses and the current location.
166 assert(DT.dominates(BB, SuccToSinkTo) &&
167 "SuccToSinkTo must be dominated by current Inst location!");
168
169 // Move the instruction.
170 Inst->moveBefore(InsertPos: SuccToSinkTo->getFirstInsertionPt());
171 return true;
172}
173
174static bool ProcessBlock(BasicBlock &BB, DominatorTree &DT, LoopInfo &LI,
175 AAResults &AA) {
176 // Don't bother sinking code out of unreachable blocks. In addition to being
177 // unprofitable, it can also lead to infinite looping, because in an
178 // unreachable loop there may be nowhere to stop.
179 if (!DT.isReachableFromEntry(A: &BB)) return false;
180
181 bool MadeChange = false;
182
183 // Walk the basic block bottom-up. Remember if we saw a store.
184 BasicBlock::iterator I = BB.end();
185 --I;
186 bool ProcessedBegin = false;
187 SmallPtrSet<Instruction *, 8> Stores;
188 do {
189 Instruction *Inst = &*I; // The instruction to sink.
190
191 // Predecrement I (if it's not begin) so that it isn't invalidated by
192 // sinking.
193 ProcessedBegin = I == BB.begin();
194 if (!ProcessedBegin)
195 --I;
196
197 if (isa<PseudoProbeInst>(Val: Inst))
198 continue;
199
200 if (SinkInstruction(Inst, Stores, DT, LI, AA)) {
201 ++NumSunk;
202 MadeChange = true;
203 }
204
205 // If we just processed the first instruction in the block, we're done.
206 } while (!ProcessedBegin);
207
208 return MadeChange;
209}
210
211static bool iterativelySinkInstructions(Function &F, DominatorTree &DT,
212 LoopInfo &LI, AAResults &AA) {
213 bool MadeChange, EverMadeChange = false;
214
215 do {
216 MadeChange = false;
217 LLVM_DEBUG(dbgs() << "Sinking iteration " << NumSinkIter << "\n");
218 // Process all basic blocks.
219 for (BasicBlock &I : F)
220 MadeChange |= ProcessBlock(BB&: I, DT, LI, AA);
221 EverMadeChange |= MadeChange;
222 NumSinkIter++;
223 } while (MadeChange);
224
225 return EverMadeChange;
226}
227
228PreservedAnalyses SinkingPass::run(Function &F, FunctionAnalysisManager &AM) {
229 auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F);
230 auto &LI = AM.getResult<LoopAnalysis>(IR&: F);
231 auto &AA = AM.getResult<AAManager>(IR&: F);
232
233 if (!iterativelySinkInstructions(F, DT, LI, AA))
234 return PreservedAnalyses::all();
235
236 PreservedAnalyses PA;
237 PA.preserveSet<CFGAnalyses>();
238 return PA;
239}
240
241namespace {
242 class SinkingLegacyPass : public FunctionPass {
243 public:
244 static char ID; // Pass identification
245 SinkingLegacyPass() : FunctionPass(ID) {
246 initializeSinkingLegacyPassPass(*PassRegistry::getPassRegistry());
247 }
248
249 bool runOnFunction(Function &F) override {
250 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
251 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
252 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
253
254 return iterativelySinkInstructions(F, DT, LI, AA);
255 }
256
257 void getAnalysisUsage(AnalysisUsage &AU) const override {
258 AU.setPreservesCFG();
259 FunctionPass::getAnalysisUsage(AU);
260 AU.addRequired<AAResultsWrapperPass>();
261 AU.addRequired<DominatorTreeWrapperPass>();
262 AU.addRequired<LoopInfoWrapperPass>();
263 }
264 };
265} // end anonymous namespace
266
267char SinkingLegacyPass::ID = 0;
268INITIALIZE_PASS_BEGIN(SinkingLegacyPass, "sink", "Code sinking", false, false)
269INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
270INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
271INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
272INITIALIZE_PASS_END(SinkingLegacyPass, "sink", "Code sinking", false, false)
273
274FunctionPass *llvm::createSinkingPass() { return new SinkingLegacyPass(); }
275