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.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AssumptionTracker.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
29 #include "llvm/IR/Constant.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/Instruction.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
39 STATISTIC(NumLoadAlignChanged,
40 "Number of loads changed by alignment assumptions");
41 STATISTIC(NumStoreAlignChanged,
42 "Number of stores changed by alignment assumptions");
43 STATISTIC(NumMemIntAlignChanged,
44 "Number of memory intrinsics changed by alignment assumptions");
47 struct AlignmentFromAssumptions : public FunctionPass {
48 static char ID; // Pass identification, replacement for typeid
49 AlignmentFromAssumptions() : FunctionPass(ID) {
50 initializeAlignmentFromAssumptionsPass(*PassRegistry::getPassRegistry());
53 bool runOnFunction(Function &F);
55 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequired<AssumptionTracker>();
57 AU.addRequired<ScalarEvolution>();
58 AU.addRequired<DominatorTreeWrapperPass>();
61 AU.addPreserved<LoopInfo>();
62 AU.addPreserved<DominatorTreeWrapperPass>();
63 AU.addPreserved<ScalarEvolution>();
66 // For memory transfers, we need a common alignment for both the source and
67 // destination. If we have a new alignment for only one operand of a transfer
68 // instruction, save it in these maps. If we reach the other operand through
69 // another assumption later, then we may change the alignment at that point.
70 DenseMap<MemTransferInst *, unsigned> NewDestAlignments, NewSrcAlignments;
72 AssumptionTracker *AT;
77 bool extractAlignmentInfo(CallInst *I, Value *&AAPtr, const SCEV *&AlignSCEV,
78 const SCEV *&OffSCEV);
79 bool processAssumption(CallInst *I);
83 char AlignmentFromAssumptions::ID = 0;
84 static const char aip_name[] = "Alignment from assumptions";
85 INITIALIZE_PASS_BEGIN(AlignmentFromAssumptions, AA_NAME,
86 aip_name, false, false)
87 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
88 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
89 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
90 INITIALIZE_PASS_END(AlignmentFromAssumptions, AA_NAME,
91 aip_name, false, false)
93 FunctionPass *llvm::createAlignmentFromAssumptionsPass() {
94 return new AlignmentFromAssumptions();
97 // Given an expression for the (constant) alignment, AlignSCEV, and an
98 // expression for the displacement between a pointer and the aligned address,
99 // DiffSCEV, compute the alignment of the displaced pointer if it can be reduced
100 // to a constant. Using SCEV to compute alignment handles the case where
101 // DiffSCEV is a recurrence with constant start such that the aligned offset
102 // is constant. e.g. {16,+,32} % 32 -> 16.
103 static unsigned getNewAlignmentDiff(const SCEV *DiffSCEV,
104 const SCEV *AlignSCEV,
105 ScalarEvolution *SE) {
106 // DiffUnits = Diff % int64_t(Alignment)
107 const SCEV *DiffAlignDiv = SE->getUDivExpr(DiffSCEV, AlignSCEV);
108 const SCEV *DiffAlign = SE->getMulExpr(DiffAlignDiv, AlignSCEV);
109 const SCEV *DiffUnitsSCEV = SE->getMinusSCEV(DiffAlign, DiffSCEV);
111 DEBUG(dbgs() << "\talignment relative to " << *AlignSCEV << " is " <<
112 *DiffUnitsSCEV << " (diff: " << *DiffSCEV << ")\n");
114 if (const SCEVConstant *ConstDUSCEV =
115 dyn_cast<SCEVConstant>(DiffUnitsSCEV)) {
116 int64_t DiffUnits = ConstDUSCEV->getValue()->getSExtValue();
118 // If the displacement is an exact multiple of the alignment, then the
119 // displaced pointer has the same alignment as the aligned pointer, so
120 // return the alignment value.
123 cast<SCEVConstant>(AlignSCEV)->getValue()->getSExtValue();
125 // If the displacement is not an exact multiple, but the remainder is a
126 // constant, then return this remainder (but only if it is a power of 2).
127 uint64_t DiffUnitsAbs = abs64(DiffUnits);
128 if (isPowerOf2_64(DiffUnitsAbs))
129 return (unsigned) DiffUnitsAbs;
135 // There is an address given by an offset OffSCEV from AASCEV which has an
136 // alignment AlignSCEV. Use that information, if possible, to compute a new
137 // alignment for Ptr.
138 static unsigned getNewAlignment(const SCEV *AASCEV, const SCEV *AlignSCEV,
139 const SCEV *OffSCEV, Value *Ptr,
140 ScalarEvolution *SE) {
141 const SCEV *PtrSCEV = SE->getSCEV(Ptr);
142 const SCEV *DiffSCEV = SE->getMinusSCEV(PtrSCEV, AASCEV);
144 // On 32-bit platforms, DiffSCEV might now have type i32 -- we've always
145 // sign-extended OffSCEV to i64, so make sure they agree again.
146 DiffSCEV = SE->getNoopOrSignExtend(DiffSCEV, OffSCEV->getType());
148 // What we really want to know is the overall offset to the aligned
149 // address. This address is displaced by the provided offset.
150 DiffSCEV = SE->getMinusSCEV(DiffSCEV, OffSCEV);
152 DEBUG(dbgs() << "AFI: alignment of " << *Ptr << " relative to " <<
153 *AlignSCEV << " and offset " << *OffSCEV <<
154 " using diff " << *DiffSCEV << "\n");
156 unsigned NewAlignment = getNewAlignmentDiff(DiffSCEV, AlignSCEV, SE);
157 DEBUG(dbgs() << "\tnew alignment: " << NewAlignment << "\n");
161 } else if (const SCEVAddRecExpr *DiffARSCEV =
162 dyn_cast<SCEVAddRecExpr>(DiffSCEV)) {
163 // The relative offset to the alignment assumption did not yield a constant,
164 // but we should try harder: if we assume that a is 32-byte aligned, then in
165 // for (i = 0; i < 1024; i += 4) r += a[i]; not all of the loads from a are
166 // 32-byte aligned, but instead alternate between 32 and 16-byte alignment.
167 // As a result, the new alignment will not be a constant, but can still
168 // be improved over the default (of 4) to 16.
170 const SCEV *DiffStartSCEV = DiffARSCEV->getStart();
171 const SCEV *DiffIncSCEV = DiffARSCEV->getStepRecurrence(*SE);
173 DEBUG(dbgs() << "\ttrying start/inc alignment using start " <<
174 *DiffStartSCEV << " and inc " << *DiffIncSCEV << "\n");
176 // Now compute the new alignment using the displacement to the value in the
177 // first iteration, and also the alignment using the per-iteration delta.
178 // If these are the same, then use that answer. Otherwise, use the smaller
179 // one, but only if it divides the larger one.
180 NewAlignment = getNewAlignmentDiff(DiffStartSCEV, AlignSCEV, SE);
181 unsigned NewIncAlignment = getNewAlignmentDiff(DiffIncSCEV, AlignSCEV, SE);
183 DEBUG(dbgs() << "\tnew start alignment: " << NewAlignment << "\n");
184 DEBUG(dbgs() << "\tnew inc alignment: " << NewIncAlignment << "\n");
186 if (!NewAlignment || !NewIncAlignment) {
188 } else if (NewAlignment > NewIncAlignment) {
189 if (NewAlignment % NewIncAlignment == 0) {
190 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
191 NewIncAlignment << "\n");
192 return NewIncAlignment;
194 } else if (NewIncAlignment > NewAlignment) {
195 if (NewIncAlignment % NewAlignment == 0) {
196 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
197 NewAlignment << "\n");
200 } else if (NewIncAlignment == NewAlignment) {
201 DEBUG(dbgs() << "\tnew start/inc alignment: " <<
202 NewAlignment << "\n");
210 bool AlignmentFromAssumptions::extractAlignmentInfo(CallInst *I,
211 Value *&AAPtr, const SCEV *&AlignSCEV,
212 const SCEV *&OffSCEV) {
213 // An alignment assume must be a statement about the least-significant
214 // bits of the pointer being zero, possibly with some offset.
215 ICmpInst *ICI = dyn_cast<ICmpInst>(I->getArgOperand(0));
219 // This must be an expression of the form: x & m == 0.
220 if (ICI->getPredicate() != ICmpInst::ICMP_EQ)
223 // Swap things around so that the RHS is 0.
224 Value *CmpLHS = ICI->getOperand(0);
225 Value *CmpRHS = ICI->getOperand(1);
226 const SCEV *CmpLHSSCEV = SE->getSCEV(CmpLHS);
227 const SCEV *CmpRHSSCEV = SE->getSCEV(CmpRHS);
228 if (CmpLHSSCEV->isZero())
229 std::swap(CmpLHS, CmpRHS);
230 else if (!CmpRHSSCEV->isZero())
233 BinaryOperator *CmpBO = dyn_cast<BinaryOperator>(CmpLHS);
234 if (!CmpBO || CmpBO->getOpcode() != Instruction::And)
237 // Swap things around so that the right operand of the and is a constant
238 // (the mask); we cannot deal with variable masks.
239 Value *AndLHS = CmpBO->getOperand(0);
240 Value *AndRHS = CmpBO->getOperand(1);
241 const SCEV *AndLHSSCEV = SE->getSCEV(AndLHS);
242 const SCEV *AndRHSSCEV = SE->getSCEV(AndRHS);
243 if (isa<SCEVConstant>(AndLHSSCEV)) {
244 std::swap(AndLHS, AndRHS);
245 std::swap(AndLHSSCEV, AndRHSSCEV);
248 const SCEVConstant *MaskSCEV = dyn_cast<SCEVConstant>(AndRHSSCEV);
252 // The mask must have some trailing ones (otherwise the condition is
253 // trivial and tells us nothing about the alignment of the left operand).
254 unsigned TrailingOnes =
255 MaskSCEV->getValue()->getValue().countTrailingOnes();
259 // Cap the alignment at the maximum with which LLVM can deal (and make sure
260 // we don't overflow the shift).
262 TrailingOnes = std::min(TrailingOnes,
263 unsigned(sizeof(unsigned) * CHAR_BIT - 1));
264 Alignment = std::min(1u << TrailingOnes, +Value::MaximumAlignment);
266 Type *Int64Ty = Type::getInt64Ty(I->getParent()->getParent()->getContext());
267 AlignSCEV = SE->getConstant(Int64Ty, Alignment);
269 // The LHS might be a ptrtoint instruction, or it might be the pointer
273 if (PtrToIntInst *PToI = dyn_cast<PtrToIntInst>(AndLHS)) {
274 AAPtr = PToI->getPointerOperand();
275 OffSCEV = SE->getConstant(Int64Ty, 0);
276 } else if (const SCEVAddExpr* AndLHSAddSCEV =
277 dyn_cast<SCEVAddExpr>(AndLHSSCEV)) {
278 // Try to find the ptrtoint; subtract it and the rest is the offset.
279 for (SCEVAddExpr::op_iterator J = AndLHSAddSCEV->op_begin(),
280 JE = AndLHSAddSCEV->op_end(); J != JE; ++J)
281 if (const SCEVUnknown *OpUnk = dyn_cast<SCEVUnknown>(*J))
282 if (PtrToIntInst *PToI = dyn_cast<PtrToIntInst>(OpUnk->getValue())) {
283 AAPtr = PToI->getPointerOperand();
284 OffSCEV = SE->getMinusSCEV(AndLHSAddSCEV, *J);
292 // Sign extend the offset to 64 bits (so that it is like all of the other
294 unsigned OffSCEVBits = OffSCEV->getType()->getPrimitiveSizeInBits();
295 if (OffSCEVBits < 64)
296 OffSCEV = SE->getSignExtendExpr(OffSCEV, Int64Ty);
297 else if (OffSCEVBits > 64)
300 AAPtr = AAPtr->stripPointerCasts();
304 bool AlignmentFromAssumptions::processAssumption(CallInst *ACall) {
306 const SCEV *AlignSCEV, *OffSCEV;
307 if (!extractAlignmentInfo(ACall, AAPtr, AlignSCEV, OffSCEV))
310 const SCEV *AASCEV = SE->getSCEV(AAPtr);
312 // Apply the assumption to all other users of the specified pointer.
313 SmallPtrSet<Instruction *, 32> Visited;
314 SmallVector<Instruction*, 16> WorkList;
315 for (User *J : AAPtr->users()) {
319 if (Instruction *K = dyn_cast<Instruction>(J))
320 if (isValidAssumeForContext(ACall, K, DL, DT))
321 WorkList.push_back(K);
324 while (!WorkList.empty()) {
325 Instruction *J = WorkList.pop_back_val();
327 if (LoadInst *LI = dyn_cast<LoadInst>(J)) {
328 unsigned NewAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
329 LI->getPointerOperand(), SE);
331 if (NewAlignment > LI->getAlignment()) {
332 LI->setAlignment(NewAlignment);
333 ++NumLoadAlignChanged;
335 } else if (StoreInst *SI = dyn_cast<StoreInst>(J)) {
336 unsigned NewAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
337 SI->getPointerOperand(), SE);
339 if (NewAlignment > SI->getAlignment()) {
340 SI->setAlignment(NewAlignment);
341 ++NumStoreAlignChanged;
343 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(J)) {
344 unsigned NewDestAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
347 // For memory transfers, we need a common alignment for both the
348 // source and destination. If we have a new alignment for this
349 // instruction, but only for one operand, save it. If we reach the
350 // other operand through another assumption later, then we may
351 // change the alignment at that point.
352 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) {
353 unsigned NewSrcAlignment = getNewAlignment(AASCEV, AlignSCEV, OffSCEV,
354 MTI->getSource(), SE);
356 DenseMap<MemTransferInst *, unsigned>::iterator DI =
357 NewDestAlignments.find(MTI);
358 unsigned AltDestAlignment = (DI == NewDestAlignments.end()) ?
361 DenseMap<MemTransferInst *, unsigned>::iterator SI =
362 NewSrcAlignments.find(MTI);
363 unsigned AltSrcAlignment = (SI == NewSrcAlignments.end()) ?
366 DEBUG(dbgs() << "\tmem trans: " << NewDestAlignment << " " <<
367 AltDestAlignment << " " << NewSrcAlignment <<
368 " " << AltSrcAlignment << "\n");
370 // Of these four alignments, pick the largest possible...
371 unsigned NewAlignment = 0;
372 if (NewDestAlignment <= std::max(NewSrcAlignment, AltSrcAlignment))
373 NewAlignment = std::max(NewAlignment, NewDestAlignment);
374 if (AltDestAlignment <= std::max(NewSrcAlignment, AltSrcAlignment))
375 NewAlignment = std::max(NewAlignment, AltDestAlignment);
376 if (NewSrcAlignment <= std::max(NewDestAlignment, AltDestAlignment))
377 NewAlignment = std::max(NewAlignment, NewSrcAlignment);
378 if (AltSrcAlignment <= std::max(NewDestAlignment, AltDestAlignment))
379 NewAlignment = std::max(NewAlignment, AltSrcAlignment);
381 if (NewAlignment > MI->getAlignment()) {
382 MI->setAlignment(ConstantInt::get(Type::getInt32Ty(
383 MI->getParent()->getContext()), NewAlignment));
384 ++NumMemIntAlignChanged;
387 NewDestAlignments.insert(std::make_pair(MTI, NewDestAlignment));
388 NewSrcAlignments.insert(std::make_pair(MTI, NewSrcAlignment));
389 } else if (NewDestAlignment > MI->getAlignment()) {
390 assert((!isa<MemIntrinsic>(MI) || isa<MemSetInst>(MI)) &&
391 "Unknown memory intrinsic");
393 MI->setAlignment(ConstantInt::get(Type::getInt32Ty(
394 MI->getParent()->getContext()), NewDestAlignment));
395 ++NumMemIntAlignChanged;
399 // Now that we've updated that use of the pointer, look for other uses of
400 // the pointer to update.
402 for (User *UJ : J->users()) {
403 Instruction *K = cast<Instruction>(UJ);
404 if (!Visited.count(K) && isValidAssumeForContext(ACall, K, DL, DT))
405 WorkList.push_back(K);
412 bool AlignmentFromAssumptions::runOnFunction(Function &F) {
413 bool Changed = false;
414 AT = &getAnalysis<AssumptionTracker>();
415 SE = &getAnalysis<ScalarEvolution>();
416 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
417 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
418 DL = DLP ? &DLP->getDataLayout() : nullptr;
420 NewDestAlignments.clear();
421 NewSrcAlignments.clear();
423 for (auto &I : AT->assumptions(&F))
424 Changed |= processAssumption(I);