1 //===---- BDCE.cpp - Bit-tracking dead code elimination -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Bit-Tracking Dead Code Elimination pass. Some
11 // instructions (shifts, some ands, ors, etc.) kill some of their input bits.
12 // We track these dead bits and remove instructions that compute only these
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DepthFirstIterator.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AssumptionCache.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CFG.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/InstIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
40 #define DEBUG_TYPE "bdce"
42 STATISTIC(NumRemoved, "Number of instructions removed (unused)");
43 STATISTIC(NumSimplified, "Number of instructions trivialized (dead bits)");
46 struct BDCE : public FunctionPass {
47 static char ID; // Pass identification, replacement for typeid
48 BDCE() : FunctionPass(ID) {
49 initializeBDCEPass(*PassRegistry::getPassRegistry());
52 bool runOnFunction(Function& F) override;
54 void getAnalysisUsage(AnalysisUsage& AU) const override {
56 AU.addRequired<AssumptionCacheTracker>();
57 AU.addRequired<DominatorTreeWrapperPass>();
60 void determineLiveOperandBits(const Instruction *UserI,
61 const Instruction *I, unsigned OperandNo,
62 const APInt &AOut, APInt &AB,
63 APInt &KnownZero, APInt &KnownOne,
64 APInt &KnownZero2, APInt &KnownOne2);
73 INITIALIZE_PASS_BEGIN(BDCE, "bdce", "Bit-Tracking Dead Code Elimination",
75 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
76 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
77 INITIALIZE_PASS_END(BDCE, "bdce", "Bit-Tracking Dead Code Elimination",
80 static bool isAlwaysLive(Instruction *I) {
81 return isa<TerminatorInst>(I) || isa<DbgInfoIntrinsic>(I) ||
82 isa<LandingPadInst>(I) || I->mayHaveSideEffects();
85 void BDCE::determineLiveOperandBits(const Instruction *UserI,
86 const Instruction *I, unsigned OperandNo,
87 const APInt &AOut, APInt &AB,
88 APInt &KnownZero, APInt &KnownOne,
89 APInt &KnownZero2, APInt &KnownOne2) {
90 unsigned BitWidth = AB.getBitWidth();
92 // We're called once per operand, but for some instructions, we need to
93 // compute known bits of both operands in order to determine the live bits of
94 // either (when both operands are instructions themselves). We don't,
95 // however, want to do this twice, so we cache the result in APInts that live
96 // in the caller. For the two-relevant-operands case, both operand values are
98 auto ComputeKnownBits = [&](unsigned BitWidth, const Value *V1,
100 KnownZero = APInt(BitWidth, 0);
101 KnownOne = APInt(BitWidth, 0);
102 computeKnownBits(const_cast<Value*>(V1), KnownZero, KnownOne, DL, 0, AC,
106 KnownZero2 = APInt(BitWidth, 0);
107 KnownOne2 = APInt(BitWidth, 0);
108 computeKnownBits(const_cast<Value*>(V2), KnownZero2, KnownOne2, DL, 0, AC,
113 switch (UserI->getOpcode()) {
115 case Instruction::Call:
116 case Instruction::Invoke:
117 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(UserI))
118 switch (II->getIntrinsicID()) {
120 case Intrinsic::bswap:
121 // The alive bits of the input are the swapped alive bits of
123 AB = AOut.byteSwap();
125 case Intrinsic::ctlz:
126 if (OperandNo == 0) {
127 // We need some output bits, so we need all bits of the
128 // input to the left of, and including, the leftmost bit
130 ComputeKnownBits(BitWidth, I, nullptr);
131 AB = APInt::getHighBitsSet(BitWidth,
132 std::min(BitWidth, KnownOne.countLeadingZeros()+1));
135 case Intrinsic::cttz:
136 if (OperandNo == 0) {
137 // We need some output bits, so we need all bits of the
138 // input to the right of, and including, the rightmost bit
140 ComputeKnownBits(BitWidth, I, nullptr);
141 AB = APInt::getLowBitsSet(BitWidth,
142 std::min(BitWidth, KnownOne.countTrailingZeros()+1));
147 case Instruction::Add:
148 case Instruction::Sub:
149 // Find the highest live output bit. We don't need any more input
150 // bits than that (adds, and thus subtracts, ripple only to the
152 AB = APInt::getLowBitsSet(BitWidth, AOut.getActiveBits());
154 case Instruction::Shl:
156 if (ConstantInt *CI =
157 dyn_cast<ConstantInt>(UserI->getOperand(1))) {
158 uint64_t ShiftAmt = CI->getLimitedValue(BitWidth-1);
159 AB = AOut.lshr(ShiftAmt);
161 // If the shift is nuw/nsw, then the high bits are not dead
162 // (because we've promised that they *must* be zero).
163 const ShlOperator *S = cast<ShlOperator>(UserI);
164 if (S->hasNoSignedWrap())
165 AB |= APInt::getHighBitsSet(BitWidth, ShiftAmt+1);
166 else if (S->hasNoUnsignedWrap())
167 AB |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
170 case Instruction::LShr:
172 if (ConstantInt *CI =
173 dyn_cast<ConstantInt>(UserI->getOperand(1))) {
174 uint64_t ShiftAmt = CI->getLimitedValue(BitWidth-1);
175 AB = AOut.shl(ShiftAmt);
177 // If the shift is exact, then the low bits are not dead
178 // (they must be zero).
179 if (cast<LShrOperator>(UserI)->isExact())
180 AB |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
183 case Instruction::AShr:
185 if (ConstantInt *CI =
186 dyn_cast<ConstantInt>(UserI->getOperand(1))) {
187 uint64_t ShiftAmt = CI->getLimitedValue(BitWidth-1);
188 AB = AOut.shl(ShiftAmt);
189 // Because the high input bit is replicated into the
190 // high-order bits of the result, if we need any of those
191 // bits, then we must keep the highest input bit.
192 if ((AOut & APInt::getHighBitsSet(BitWidth, ShiftAmt))
194 AB.setBit(BitWidth-1);
196 // If the shift is exact, then the low bits are not dead
197 // (they must be zero).
198 if (cast<AShrOperator>(UserI)->isExact())
199 AB |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
202 case Instruction::And:
205 // For bits that are known zero, the corresponding bits in the
206 // other operand are dead (unless they're both zero, in which
207 // case they can't both be dead, so just mark the LHS bits as
209 if (OperandNo == 0) {
210 ComputeKnownBits(BitWidth, I, UserI->getOperand(1));
213 if (!isa<Instruction>(UserI->getOperand(0)))
214 ComputeKnownBits(BitWidth, UserI->getOperand(0), I);
215 AB &= ~(KnownZero & ~KnownZero2);
218 case Instruction::Or:
221 // For bits that are known one, the corresponding bits in the
222 // other operand are dead (unless they're both one, in which
223 // case they can't both be dead, so just mark the LHS bits as
225 if (OperandNo == 0) {
226 ComputeKnownBits(BitWidth, I, UserI->getOperand(1));
229 if (!isa<Instruction>(UserI->getOperand(0)))
230 ComputeKnownBits(BitWidth, UserI->getOperand(0), I);
231 AB &= ~(KnownOne & ~KnownOne2);
234 case Instruction::Xor:
235 case Instruction::PHI:
238 case Instruction::Trunc:
239 AB = AOut.zext(BitWidth);
241 case Instruction::ZExt:
242 AB = AOut.trunc(BitWidth);
244 case Instruction::SExt:
245 AB = AOut.trunc(BitWidth);
246 // Because the high input bit is replicated into the
247 // high-order bits of the result, if we need any of those
248 // bits, then we must keep the highest input bit.
249 if ((AOut & APInt::getHighBitsSet(AOut.getBitWidth(),
250 AOut.getBitWidth() - BitWidth))
252 AB.setBit(BitWidth-1);
254 case Instruction::Select:
261 bool BDCE::runOnFunction(Function& F) {
262 if (skipOptnoneFunction(F))
265 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
266 DL = F.getParent()->getDataLayout();
267 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
269 DenseMap<Instruction *, APInt> AliveBits;
270 SmallVector<Instruction*, 128> Worklist;
272 // The set of visited instructions (non-integer-typed only).
273 SmallPtrSet<Instruction*, 128> Visited;
275 // Collect the set of "root" instructions that are known live.
276 for (Instruction &I : inst_range(F)) {
277 if (!isAlwaysLive(&I))
280 DEBUG(dbgs() << "BDCE: Root: " << I);
281 // For integer-valued instructions, set up an initial empty set of alive
282 // bits and add the instruction to the work list. For other instructions
283 // add their operands to the work list (for integer values operands, mark
284 // all bits as live).
285 if (IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
286 if (!AliveBits.count(&I)) {
287 AliveBits[&I] = APInt(IT->getBitWidth(), 0);
288 Worklist.push_back(&I);
294 // Non-integer-typed instructions...
295 for (Use &OI : I.operands()) {
296 if (Instruction *J = dyn_cast<Instruction>(OI)) {
297 if (IntegerType *IT = dyn_cast<IntegerType>(J->getType()))
298 AliveBits[J] = APInt::getAllOnesValue(IT->getBitWidth());
299 Worklist.push_back(J);
302 // To save memory, we don't add I to the Visited set here. Instead, we
303 // check isAlwaysLive on every instruction when searching for dead
304 // instructions later (we need to check isAlwaysLive for the
305 // integer-typed instructions anyway).
308 // Propagate liveness backwards to operands.
309 while (!Worklist.empty()) {
310 Instruction *UserI = Worklist.pop_back_val();
312 DEBUG(dbgs() << "BDCE: Visiting: " << *UserI);
314 if (UserI->getType()->isIntegerTy()) {
315 AOut = AliveBits[UserI];
316 DEBUG(dbgs() << " Alive Out: " << AOut);
318 DEBUG(dbgs() << "\n");
320 if (!UserI->getType()->isIntegerTy())
321 Visited.insert(UserI);
323 APInt KnownZero, KnownOne, KnownZero2, KnownOne2;
324 // Compute the set of alive bits for each operand. These are anded into the
325 // existing set, if any, and if that changes the set of alive bits, the
326 // operand is added to the work-list.
327 for (Use &OI : UserI->operands()) {
328 if (Instruction *I = dyn_cast<Instruction>(OI)) {
329 if (IntegerType *IT = dyn_cast<IntegerType>(I->getType())) {
330 unsigned BitWidth = IT->getBitWidth();
331 APInt AB = APInt::getAllOnesValue(BitWidth);
332 if (UserI->getType()->isIntegerTy() && !AOut &&
333 !isAlwaysLive(UserI)) {
334 AB = APInt(BitWidth, 0);
336 // If all bits of the output are dead, then all bits of the input
337 // Bits of each operand that are used to compute alive bits of the
338 // output are alive, all others are dead.
339 determineLiveOperandBits(UserI, I, OI.getOperandNo(), AOut, AB,
341 KnownZero2, KnownOne2);
344 // If we've added to the set of alive bits (or the operand has not
345 // been previously visited), then re-queue the operand to be visited
347 APInt ABPrev(BitWidth, 0);
348 auto ABI = AliveBits.find(I);
349 if (ABI != AliveBits.end())
350 ABPrev = ABI->second;
352 APInt ABNew = AB | ABPrev;
353 if (ABNew != ABPrev || ABI == AliveBits.end()) {
354 AliveBits[I] = std::move(ABNew);
355 Worklist.push_back(I);
357 } else if (!Visited.count(I)) {
358 Worklist.push_back(I);
364 bool Changed = false;
365 // The inverse of the live set is the dead set. These are those instructions
366 // which have no side effects and do not influence the control flow or return
367 // value of the function, and may therefore be deleted safely.
368 // NOTE: We reuse the Worklist vector here for memory efficiency.
369 for (Instruction &I : inst_range(F)) {
370 // For live instructions that have all dead bits, first make them dead by
371 // replacing all uses with something else. Then, if they don't need to
372 // remain live (because they have side effects, etc.) we can remove them.
373 if (I.getType()->isIntegerTy()) {
374 auto ABI = AliveBits.find(&I);
375 if (ABI != AliveBits.end()) {
376 if (ABI->second.getBoolValue())
379 DEBUG(dbgs() << "BDCE: Trivializing: " << I << " (all bits dead)\n");
380 // FIXME: In theory we could substitute undef here instead of zero.
381 // This should be reconsidered once we settle on the semantics of
382 // undef, poison, etc.
383 Value *Zero = ConstantInt::get(I.getType(), 0);
385 I.replaceAllUsesWith(Zero);
388 } else if (Visited.count(&I)) {
392 if (isAlwaysLive(&I))
395 Worklist.push_back(&I);
396 I.dropAllReferences();
400 for (Instruction *&I : Worklist) {
402 I->eraseFromParent();
408 FunctionPass *llvm::createBitTrackingDCEPass() {