1 //===----------------------- AlignmentFromAssumptions.cpp -----------------===//
2 // Set Load/Store Alignments From Assumptions
4 // The LLVM Compiler Infrastructure
6 // This file is distributed under the University of Illinois Open Source
7 // License. See LICENSE.TXT for details.
9 //===----------------------------------------------------------------------===//
11 // This file implements a ScalarEvolution-based transformation to set
12 // the alignments of load, stores and memory intrinsics based on the truth
13 // expressions of assume intrinsics. The primary motivation is to handle
14 // complex alignment assumptions that apply to vector loads and stores that
15 // appear after vectorization and unrolling.
17 //===----------------------------------------------------------------------===//
19 #define AA_NAME "alignment-from-assumptions"
20 #define DEBUG_TYPE AA_NAME
21 #include "llvm/Transforms/Scalar/AlignmentFromAssumptions.h"
22 #include "llvm/Transforms/Scalar.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/GlobalsModRef.h"
27 #include "llvm/Analysis/AssumptionCache.h"
28 #include "llvm/Analysis/LoopInfo.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/Constant.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Instruction.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/raw_ostream.h"
40 STATISTIC(NumLoadAlignChanged,
41 "Number of loads changed by alignment assumptions");
42 STATISTIC(NumStoreAlignChanged,
43 "Number of stores changed by alignment assumptions");
44 STATISTIC(NumMemIntAlignChanged,
45 "Number of memory intrinsics changed by alignment assumptions");
48 struct AlignmentFromAssumptions : public FunctionPass {
49 static char ID; // Pass identification, replacement for typeid
50 AlignmentFromAssumptions() : FunctionPass(ID) {
51 initializeAlignmentFromAssumptionsPass(*PassRegistry::getPassRegistry());
54 bool runOnFunction(Function &F) override;
56 void getAnalysisUsage(AnalysisUsage &AU) const override {
57 AU.addRequired<AssumptionCacheTracker>();
58 AU.addRequired<ScalarEvolutionWrapperPass>();
59 AU.addRequired<DominatorTreeWrapperPass>();
62 AU.addPreserved<AAResultsWrapperPass>();
63 AU.addPreserved<GlobalsAAWrapperPass>();
64 AU.addPreserved<LoopInfoWrapperPass>();
65 AU.addPreserved<DominatorTreeWrapperPass>();
66 AU.addPreserved<ScalarEvolutionWrapperPass>();
69 AlignmentFromAssumptionsPass Impl;
73 char AlignmentFromAssumptions::ID = 0;
74 static const char aip_name[] = "Alignment from assumptions";
75 INITIALIZE_PASS_BEGIN(AlignmentFromAssumptions, AA_NAME,
76 aip_name, false, false)
77 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
78 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
79 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
80 INITIALIZE_PASS_END(AlignmentFromAssumptions, AA_NAME,
81 aip_name, false, false)
83 FunctionPass *llvm::createAlignmentFromAssumptionsPass() {
84 return new AlignmentFromAssumptions();
87 // Given an expression for the (constant) alignment, AlignSCEV, and an
88 // expression for the displacement between a pointer and the aligned address,
89 // DiffSCEV, compute the alignment of the displaced pointer if it can be reduced
90 // to a constant. Using SCEV to compute alignment handles the case where
91 // DiffSCEV is a recurrence with constant start such that the aligned offset
92 // is constant. e.g. {16,+,32} % 32 -> 16.
93 static unsigned getNewAlignmentDiff(const SCEV *DiffSCEV,
94 const SCEV *AlignSCEV,
95 ScalarEvolution *SE) {
96 // DiffUnits = Diff % int64_t(Alignment)
97 const SCEV *DiffAlignDiv = SE->getUDivExpr(DiffSCEV, AlignSCEV);
98 const SCEV *DiffAlign = SE->getMulExpr(DiffAlignDiv, AlignSCEV);
99 const SCEV *DiffUnitsSCEV = SE->getMinusSCEV(DiffAlign, DiffSCEV);
101 DEBUG(dbgs() << "\talignment relative to " << *AlignSCEV << " is " <<
102 *DiffUnitsSCEV << " (diff: " << *DiffSCEV << ")\n");
104 if (const SCEVConstant *ConstDUSCEV =
105 dyn_cast<SCEVConstant>(DiffUnitsSCEV)) {
106 int64_t DiffUnits = ConstDUSCEV->getValue()->getSExtValue();
108 // If the displacement is an exact multiple of the alignment, then the
109 // displaced pointer has the same alignment as the aligned pointer, so
110 // return the alignment value.
113 cast<SCEVConstant>(AlignSCEV)->getValue()->getSExtValue();
115 // If the displacement is not an exact multiple, but the remainder is a
116 // constant, then return this remainder (but only if it is a power of 2).
117 uint64_t DiffUnitsAbs = std::abs(DiffUnits);
118 if (isPowerOf2_64(DiffUnitsAbs))
119 return (unsigned) DiffUnitsAbs;
125 // There is an address given by an offset OffSCEV from AASCEV which has an
126 // alignment AlignSCEV. Use that information, if possible, to compute a new
127 // alignment for Ptr.
128 static unsigned getNewAlignment(const SCEV *AASCEV, const SCEV *AlignSCEV,
129 const SCEV *OffSCEV, Value *Ptr,
130 ScalarEvolution *SE) {
131 const SCEV *PtrSCEV = SE->getSCEV(Ptr);
132 const SCEV *DiffSCEV = SE->getMinusSCEV(PtrSCEV, AASCEV);
134 // On 32-bit platforms, DiffSCEV might now have type i32 -- we've always
135 // sign-extended OffSCEV to i64, so make sure they agree again.
136 DiffSCEV = SE->getNoopOrSignExtend(DiffSCEV, OffSCEV->getType());
138 // What we really want to know is the overall offset to the aligned
139 // address. This address is displaced by the provided offset.
140 DiffSCEV = SE->getMinusSCEV(DiffSCEV, OffSCEV);
142 DEBUG(dbgs() << "AFI: alignment of " << *Ptr << " relative to " <<
143 *AlignSCEV << " and offset " << *OffSCEV <<
144 " using diff " << *DiffSCEV << "\n");
146 unsigned NewAlignment = getNewAlignmentDiff(DiffSCEV, AlignSCEV, SE);
147 DEBUG(dbgs() << "\tnew alignment: " << NewAlignment << "\n");
151 } else if (const SCEVAddRecExpr *DiffARSCEV =
152 dyn_cast<SCEVAddRecExpr>(DiffSCEV)) {
153 // The relative offset to the alignment assumption did not yield a constant,
154 // but we should try harder: if we assume that a is 32-byte aligned, then in
155 // for (i = 0; i < 1024; i += 4) r += a[i]; not all of the loads from a are
156 // 32-byte aligned, but instead alternate between 32 and 16-byte alignment.
157 // As a result, the new alignment will not be a constant, but can still
158 // be improved over the default (of 4) to 16.
160 const SCEV *DiffStartSCEV = DiffARSCEV->getStart();
161 const SCEV *DiffIncSCEV = DiffARSCEV->getStepRecurrence(*SE);
163 DEBUG(dbgs() << "\ttrying start/inc alignment using start " <<
164 *DiffStartSCEV << " and inc " << *DiffIncSCEV << "\n");
166 // Now compute the new alignment using the displacement to the value in the
167 // first iteration, and also the alignment using the per-iteration delta.
168 // If these are the same, then use that answer. Otherwise, use the smaller
169 // one, but only if it divides the larger one.
170 NewAlignment = getNewAlignmentDiff(DiffStartSCEV, AlignSCEV, SE);
171 unsigned NewIncAlignment = getNewAlignmentDiff(DiffIncSCEV, AlignSCEV, SE);
173 DEBUG(dbgs() << "\tnew start alignment: " << NewAlignment << "\n");
174 DEBUG(dbgs() << "\tnew inc alignment: " << NewIncAlignment << "\n");
176 if (!NewAlignment || !NewIncAlignment) {
178 } else if (NewAlignment > NewIncAlignment) {
179 if (NewAlignment % NewIncAlignment == 0) {
180 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
181 NewIncAlignment << "\n");
182 return NewIncAlignment;
184 } else if (NewIncAlignment > NewAlignment) {
185 if (NewIncAlignment % NewAlignment == 0) {
186 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
187 NewAlignment << "\n");
190 } else if (NewIncAlignment == NewAlignment) {
191 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
192 NewAlignment << "\n");
200 bool AlignmentFromAssumptionsPass::extractAlignmentInfo(CallInst *I,
202 const SCEV *&AlignSCEV,
203 const SCEV *&OffSCEV) {
204 // An alignment assume must be a statement about the least-significant
205 // bits of the pointer being zero, possibly with some offset.
206 ICmpInst *ICI = dyn_cast<ICmpInst>(I->getArgOperand(0));
210 // This must be an expression of the form: x & m == 0.
211 if (ICI->getPredicate() != ICmpInst::ICMP_EQ)
214 // Swap things around so that the RHS is 0.
215 Value *CmpLHS = ICI->getOperand(0);
216 Value *CmpRHS = ICI->getOperand(1);
217 const SCEV *CmpLHSSCEV = SE->getSCEV(CmpLHS);
218 const SCEV *CmpRHSSCEV = SE->getSCEV(CmpRHS);
219 if (CmpLHSSCEV->isZero())
220 std::swap(CmpLHS, CmpRHS);
221 else if (!CmpRHSSCEV->isZero())
224 BinaryOperator *CmpBO = dyn_cast<BinaryOperator>(CmpLHS);
225 if (!CmpBO || CmpBO->getOpcode() != Instruction::And)
228 // Swap things around so that the right operand of the and is a constant
229 // (the mask); we cannot deal with variable masks.
230 Value *AndLHS = CmpBO->getOperand(0);
231 Value *AndRHS = CmpBO->getOperand(1);
232 const SCEV *AndLHSSCEV = SE->getSCEV(AndLHS);
233 const SCEV *AndRHSSCEV = SE->getSCEV(AndRHS);
234 if (isa<SCEVConstant>(AndLHSSCEV)) {
235 std::swap(AndLHS, AndRHS);
236 std::swap(AndLHSSCEV, AndRHSSCEV);
239 const SCEVConstant *MaskSCEV = dyn_cast<SCEVConstant>(AndRHSSCEV);
243 // The mask must have some trailing ones (otherwise the condition is
244 // trivial and tells us nothing about the alignment of the left operand).
245 unsigned TrailingOnes = MaskSCEV->getAPInt().countTrailingOnes();
249 // Cap the alignment at the maximum with which LLVM can deal (and make sure
250 // we don't overflow the shift).
252 TrailingOnes = std::min(TrailingOnes,
253 unsigned(sizeof(unsigned) * CHAR_BIT - 1));
254 Alignment = std::min(1u << TrailingOnes, +Value::MaximumAlignment);
256 Type *Int64Ty = Type::getInt64Ty(I->getParent()->getParent()->getContext());
257 AlignSCEV = SE->getConstant(Int64Ty, Alignment);
259 // The LHS might be a ptrtoint instruction, or it might be the pointer
263 if (PtrToIntInst *PToI = dyn_cast<PtrToIntInst>(AndLHS)) {
264 AAPtr = PToI->getPointerOperand();
265 OffSCEV = SE->getZero(Int64Ty);
266 } else if (const SCEVAddExpr* AndLHSAddSCEV =
267 dyn_cast<SCEVAddExpr>(AndLHSSCEV)) {
268 // Try to find the ptrtoint; subtract it and the rest is the offset.
269 for (SCEVAddExpr::op_iterator J = AndLHSAddSCEV->op_begin(),
270 JE = AndLHSAddSCEV->op_end(); J != JE; ++J)
271 if (const SCEVUnknown *OpUnk = dyn_cast<SCEVUnknown>(*J))
272 if (PtrToIntInst *PToI = dyn_cast<PtrToIntInst>(OpUnk->getValue())) {
273 AAPtr = PToI->getPointerOperand();
274 OffSCEV = SE->getMinusSCEV(AndLHSAddSCEV, *J);
282 // Sign extend the offset to 64 bits (so that it is like all of the other
284 unsigned OffSCEVBits = OffSCEV->getType()->getPrimitiveSizeInBits();
285 if (OffSCEVBits < 64)
286 OffSCEV = SE->getSignExtendExpr(OffSCEV, Int64Ty);
287 else if (OffSCEVBits > 64)
290 AAPtr = AAPtr->stripPointerCasts();
294 bool AlignmentFromAssumptionsPass::processAssumption(CallInst *ACall) {
296 const SCEV *AlignSCEV, *OffSCEV;
297 if (!extractAlignmentInfo(ACall, AAPtr, AlignSCEV, OffSCEV))
300 // Skip ConstantPointerNull and UndefValue. Assumptions on these shouldn't
301 // affect other users.
302 if (isa<ConstantData>(AAPtr))
305 const SCEV *AASCEV = SE->getSCEV(AAPtr);
307 // Apply the assumption to all other users of the specified pointer.
308 SmallPtrSet<Instruction *, 32> Visited;
309 SmallVector<Instruction*, 16> WorkList;
310 for (User *J : AAPtr->users()) {
314 if (Instruction *K = dyn_cast<Instruction>(J))
315 if (isValidAssumeForContext(ACall, K, DT))
316 WorkList.push_back(K);
319 while (!WorkList.empty()) {
320 Instruction *J = WorkList.pop_back_val();
322 if (LoadInst *LI = dyn_cast<LoadInst>(J)) {
323 unsigned NewAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
324 LI->getPointerOperand(), SE);
326 if (NewAlignment > LI->getAlignment()) {
327 LI->setAlignment(NewAlignment);
328 ++NumLoadAlignChanged;
330 } else if (StoreInst *SI = dyn_cast<StoreInst>(J)) {
331 unsigned NewAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
332 SI->getPointerOperand(), SE);
334 if (NewAlignment > SI->getAlignment()) {
335 SI->setAlignment(NewAlignment);
336 ++NumStoreAlignChanged;
338 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(J)) {
339 unsigned NewDestAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
342 // For memory transfers, we need a common alignment for both the
343 // source and destination. If we have a new alignment for this
344 // instruction, but only for one operand, save it. If we reach the
345 // other operand through another assumption later, then we may
346 // change the alignment at that point.
347 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
348 unsigned NewSrcAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
349 MTI->getSource(), SE);
351 DenseMap<MemTransferInst *, unsigned>::iterator DI =
352 NewDestAlignments.find(MTI);
353 unsigned AltDestAlignment = (DI == NewDestAlignments.end()) ?
356 DenseMap<MemTransferInst *, unsigned>::iterator SI =
357 NewSrcAlignments.find(MTI);
358 unsigned AltSrcAlignment = (SI == NewSrcAlignments.end()) ?
361 DEBUG(dbgs() << "\tmem trans: " << NewDestAlignment << " " <<
362 AltDestAlignment << " " << NewSrcAlignment <<
363 " " << AltSrcAlignment << "\n");
365 // Of these four alignments, pick the largest possible...
366 unsigned NewAlignment = 0;
367 if (NewDestAlignment <= std::max(NewSrcAlignment, AltSrcAlignment))
368 NewAlignment = std::max(NewAlignment, NewDestAlignment);
369 if (AltDestAlignment <= std::max(NewSrcAlignment, AltSrcAlignment))
370 NewAlignment = std::max(NewAlignment, AltDestAlignment);
371 if (NewSrcAlignment <= std::max(NewDestAlignment, AltDestAlignment))
372 NewAlignment = std::max(NewAlignment, NewSrcAlignment);
373 if (AltSrcAlignment <= std::max(NewDestAlignment, AltDestAlignment))
374 NewAlignment = std::max(NewAlignment, AltSrcAlignment);
376 if (NewAlignment > MI->getAlignment()) {
377 MI->setAlignment(ConstantInt::get(Type::getInt32Ty(
378 MI->getParent()->getContext()), NewAlignment));
379 ++NumMemIntAlignChanged;
382 NewDestAlignments.insert(std::make_pair(MTI, NewDestAlignment));
383 NewSrcAlignments.insert(std::make_pair(MTI, NewSrcAlignment));
384 } else if (NewDestAlignment > MI->getAlignment()) {
385 assert((!isa<MemIntrinsic>(MI) || isa<MemSetInst>(MI)) &&
386 "Unknown memory intrinsic");
388 MI->setAlignment(ConstantInt::get(Type::getInt32Ty(
389 MI->getParent()->getContext()), NewDestAlignment));
390 ++NumMemIntAlignChanged;
394 // Now that we've updated that use of the pointer, look for other uses of
395 // the pointer to update.
397 for (User *UJ : J->users()) {
398 Instruction *K = cast<Instruction>(UJ);
399 if (!Visited.count(K) && isValidAssumeForContext(ACall, K, DT))
400 WorkList.push_back(K);
407 bool AlignmentFromAssumptions::runOnFunction(Function &F) {
411 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
412 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
413 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
415 return Impl.runImpl(F, AC, SE, DT);
418 bool AlignmentFromAssumptionsPass::runImpl(Function &F, AssumptionCache &AC,
419 ScalarEvolution *SE_,
420 DominatorTree *DT_) {
424 NewDestAlignments.clear();
425 NewSrcAlignments.clear();
427 bool Changed = false;
428 for (auto &AssumeVH : AC.assumptions())
430 Changed |= processAssumption(cast<CallInst>(AssumeVH));
436 AlignmentFromAssumptionsPass::run(Function &F, FunctionAnalysisManager &AM) {
438 AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(F);
439 ScalarEvolution &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
440 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
441 bool Changed = runImpl(F, AC, &SE, &DT);
443 // FIXME: We need to invalidate this to avoid PR28400. Is there a better
445 AM.invalidate<ScalarEvolutionAnalysis>(F);
448 return PreservedAnalyses::all();
449 PreservedAnalyses PA;
450 PA.preserve<AAManager>();
451 PA.preserve<ScalarEvolutionAnalysis>();
452 PA.preserve<GlobalsAA>();
453 PA.preserve<LoopAnalysis>();
454 PA.preserve<DominatorTreeAnalysis>();