Line data Source code
1 : //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis ------------===//
2 : //
3 : // The LLVM Compiler Infrastructure
4 : //
5 : // This file is distributed under the University of Illinois Open Source
6 : // License. See LICENSE.TXT for details.
7 : //
8 : //===----------------------------------------------------------------------===//
9 : //
10 : // This file contains the implementation of the scalar evolution expander,
11 : // which is used to generate the code corresponding to a given scalar evolution
12 : // expression.
13 : //
14 : //===----------------------------------------------------------------------===//
15 :
16 : #include "llvm/Analysis/ScalarEvolutionExpander.h"
17 : #include "llvm/ADT/STLExtras.h"
18 : #include "llvm/ADT/SmallSet.h"
19 : #include "llvm/Analysis/InstructionSimplify.h"
20 : #include "llvm/Analysis/LoopInfo.h"
21 : #include "llvm/Analysis/TargetTransformInfo.h"
22 : #include "llvm/IR/DataLayout.h"
23 : #include "llvm/IR/Dominators.h"
24 : #include "llvm/IR/IntrinsicInst.h"
25 : #include "llvm/IR/LLVMContext.h"
26 : #include "llvm/IR/Module.h"
27 : #include "llvm/IR/PatternMatch.h"
28 : #include "llvm/Support/Debug.h"
29 : #include "llvm/Support/raw_ostream.h"
30 :
31 : using namespace llvm;
32 : using namespace PatternMatch;
33 :
34 : /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
35 : /// reusing an existing cast if a suitable one exists, moving an existing
36 : /// cast if a suitable one exists but isn't in the right place, or
37 : /// creating a new one.
38 10266 : Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty,
39 : Instruction::CastOps Op,
40 : BasicBlock::iterator IP) {
41 : // This function must be called with the builder having a valid insertion
42 : // point. It doesn't need to be the actual IP where the uses of the returned
43 : // cast will be added, but it must dominate such IP.
44 : // We use this precondition to produce a cast that will dominate all its
45 : // uses. In particular, this is crucial for the case where the builder's
46 : // insertion point *is* the point where we were asked to put the cast.
47 : // Since we don't know the builder's insertion point is actually
48 : // where the uses will be added (only that it dominates it), we are
49 : // not allowed to move it.
50 : BasicBlock::iterator BIP = Builder.GetInsertPoint();
51 :
52 : Instruction *Ret = nullptr;
53 :
54 : // Check to see if there is already a cast!
55 20484 : for (User *U : V->users())
56 12583 : if (U->getType() == Ty)
57 : if (CastInst *CI = dyn_cast<CastInst>(U))
58 2365 : if (CI->getOpcode() == Op) {
59 : // If the cast isn't where we want it, create a new cast at IP.
60 : // Likewise, do not reuse a cast at BIP because it must dominate
61 : // instructions that might be inserted before BIP.
62 2365 : if (BasicBlock::iterator(CI) != IP || BIP == IP) {
63 : // Create a new cast, and leave the old cast in place in case
64 : // it is being used as an insert point. Clear its operand
65 : // so that it doesn't hold anything live.
66 374 : Ret = CastInst::Create(Op, V, Ty, "", &*IP);
67 374 : Ret->takeName(CI);
68 374 : CI->replaceAllUsesWith(Ret);
69 374 : CI->setOperand(0, UndefValue::get(V->getType()));
70 : break;
71 : }
72 : Ret = CI;
73 : break;
74 : }
75 :
76 : // Create a new cast.
77 10266 : if (!Ret)
78 7901 : Ret = CastInst::Create(Op, V, Ty, V->getName(), &*IP);
79 :
80 : // We assert at the end of the function since IP might point to an
81 : // instruction with different dominance properties than a cast
82 : // (an invoke for example) and not dominate BIP (but the cast does).
83 : assert(SE.DT.dominates(Ret, &*BIP));
84 :
85 10266 : rememberInstruction(Ret);
86 10266 : return Ret;
87 : }
88 :
89 9074 : static BasicBlock::iterator findInsertPointAfter(Instruction *I,
90 : BasicBlock *MustDominate) {
91 9074 : BasicBlock::iterator IP = ++I->getIterator();
92 : if (auto *II = dyn_cast<InvokeInst>(I))
93 : IP = II->getNormalDest()->begin();
94 :
95 13814 : while (isa<PHINode>(IP))
96 : ++IP;
97 :
98 9074 : if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) {
99 : ++IP;
100 9073 : } else if (isa<CatchSwitchInst>(IP)) {
101 : IP = MustDominate->getFirstInsertionPt();
102 : } else {
103 : assert(!IP->isEHPad() && "unexpected eh pad!");
104 : }
105 :
106 9074 : return IP;
107 : }
108 :
109 : /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
110 : /// which must be possible with a noop cast, doing what we can to share
111 : /// the casts.
112 87248 : Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) {
113 87248 : Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
114 : assert((Op == Instruction::BitCast ||
115 : Op == Instruction::PtrToInt ||
116 : Op == Instruction::IntToPtr) &&
117 : "InsertNoopCastOfTo cannot perform non-noop casts!");
118 : assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
119 : "InsertNoopCastOfTo cannot change sizes!");
120 :
121 : // Short-circuit unnecessary bitcasts.
122 87248 : if (Op == Instruction::BitCast) {
123 86125 : if (V->getType() == Ty)
124 : return V;
125 : if (CastInst *CI = dyn_cast<CastInst>(V)) {
126 106 : if (CI->getOperand(0)->getType() == Ty)
127 : return CI->getOperand(0);
128 : }
129 : }
130 : // Short-circuit unnecessary inttoptr<->ptrtoint casts.
131 12164 : if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
132 1123 : SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
133 : if (CastInst *CI = dyn_cast<CastInst>(V))
134 174 : if ((CI->getOpcode() == Instruction::PtrToInt ||
135 351 : CI->getOpcode() == Instruction::IntToPtr) &&
136 175 : SE.getTypeSizeInBits(CI->getType()) ==
137 175 : SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
138 : return CI->getOperand(0);
139 : if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
140 10 : if ((CE->getOpcode() == Instruction::PtrToInt ||
141 10 : CE->getOpcode() == Instruction::IntToPtr) &&
142 0 : SE.getTypeSizeInBits(CE->getType()) ==
143 0 : SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
144 : return CE->getOperand(0);
145 : }
146 :
147 : // Fold a cast of a constant.
148 : if (Constant *C = dyn_cast<Constant>(V))
149 600 : return ConstantExpr::getCast(Op, C, Ty);
150 :
151 : // Cast the argument at the beginning of the entry block, after
152 : // any bitcasts of other arguments.
153 : if (Argument *A = dyn_cast<Argument>(V)) {
154 2406 : BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
155 1173 : while ((isa<BitCastInst>(IP) &&
156 1172 : isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
157 3920 : cast<BitCastInst>(IP)->getOperand(0) != A) ||
158 : isa<DbgInfoIntrinsic>(IP))
159 : ++IP;
160 1203 : return ReuseOrCreateCast(A, Ty, Op, IP);
161 : }
162 :
163 : // Cast the instruction immediately after the instruction.
164 : Instruction *I = cast<Instruction>(V);
165 9063 : BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock());
166 9063 : return ReuseOrCreateCast(I, Ty, Op, IP);
167 : }
168 :
169 : /// InsertBinop - Insert the specified binary operator, doing a small amount
170 : /// of work to avoid inserting an obviously redundant operation.
171 7249 : Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
172 : Value *LHS, Value *RHS) {
173 : // Fold a binop with constant operands.
174 : if (Constant *CLHS = dyn_cast<Constant>(LHS))
175 : if (Constant *CRHS = dyn_cast<Constant>(RHS))
176 223 : return ConstantExpr::get(Opcode, CLHS, CRHS);
177 :
178 : // Do a quick scan to see if we have this binop nearby. If so, reuse it.
179 : unsigned ScanLimit = 6;
180 7026 : BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
181 : // Scanning starts from the last instruction before the insertion point.
182 7026 : BasicBlock::iterator IP = Builder.GetInsertPoint();
183 7026 : if (IP != BlockBegin) {
184 : --IP;
185 27651 : for (; ScanLimit; --IP, --ScanLimit) {
186 : // Don't count dbg.value against the ScanLimit, to avoid perturbing the
187 : // generated code.
188 25595 : if (isa<DbgInfoIntrinsic>(IP))
189 890 : ScanLimit++;
190 :
191 : // Conservatively, do not use any instruction which has any of wrap/exact
192 : // flags installed.
193 : // TODO: Instead of simply disable poison instructions we can be clever
194 : // here and match SCEV to this instruction.
195 : auto canGeneratePoison = [](Instruction *I) {
196 : if (isa<OverflowingBinaryOperator>(I) &&
197 : (I->hasNoSignedWrap() || I->hasNoUnsignedWrap()))
198 : return true;
199 : if (isa<PossiblyExactOperator>(I) && I->isExact())
200 : return true;
201 : return false;
202 : };
203 11252 : if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
204 25963 : IP->getOperand(1) == RHS && !canGeneratePoison(&*IP))
205 : return &*IP;
206 25335 : if (IP == BlockBegin) break;
207 : }
208 : }
209 :
210 : // Save the original insertion point so we can restore it when we're done.
211 : DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc();
212 13532 : SCEVInsertPointGuard Guard(Builder, this);
213 :
214 : // Move the insertion point out of as many loops as we can.
215 8820 : while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
216 2008 : if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
217 61 : BasicBlock *Preheader = L->getLoopPreheader();
218 61 : if (!Preheader) break;
219 :
220 : // Ok, move up a level.
221 46 : Builder.SetInsertPoint(Preheader->getTerminator());
222 46 : }
223 :
224 : // If we haven't found this binop, insert it.
225 13532 : Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS));
226 6766 : BO->setDebugLoc(Loc);
227 6766 : rememberInstruction(BO);
228 :
229 : return BO;
230 : }
231 :
232 : /// FactorOutConstant - Test if S is divisible by Factor, using signed
233 : /// division. If so, update S with Factor divided out and return true.
234 : /// S need not be evenly divisible if a reasonable remainder can be
235 : /// computed.
236 : /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
237 : /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
238 : /// check to see if the divide was folded.
239 18723 : static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder,
240 : const SCEV *Factor, ScalarEvolution &SE,
241 : const DataLayout &DL) {
242 : // Everything is divisible by one.
243 18723 : if (Factor->isOne())
244 : return true;
245 :
246 : // x/x == 1.
247 15047 : if (S == Factor) {
248 1458 : S = SE.getConstant(S->getType(), 1);
249 1458 : return true;
250 : }
251 :
252 : // For a Constant, check for a multiple of the given factor.
253 : if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
254 : // 0/x == 0.
255 7017 : if (C->isZero())
256 : return true;
257 : // Check for divisibility.
258 : if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
259 : ConstantInt *CI =
260 13768 : ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt()));
261 : // If the quotient is zero and the remainder is non-zero, reject
262 : // the value at this scale. It will be considered for subsequent
263 : // smaller scales.
264 6884 : if (!CI->isZero()) {
265 5446 : const SCEV *Div = SE.getConstant(CI);
266 5446 : S = Div;
267 5446 : Remainder = SE.getAddExpr(
268 5446 : Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt())));
269 5446 : return true;
270 : }
271 : }
272 : }
273 :
274 : // In a Mul, check if there is a constant operand which is a multiple
275 : // of the given factor.
276 8010 : if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
277 : // Size is known, check if there is a constant operand which is a multiple
278 : // of the given factor. If so, we can factor it.
279 : const SCEVConstant *FC = cast<SCEVConstant>(Factor);
280 4632 : if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
281 9256 : if (!C->getAPInt().srem(FC->getAPInt())) {
282 3346 : SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
283 3346 : NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt()));
284 3346 : S = SE.getMulExpr(NewMulOps);
285 : return true;
286 : }
287 : }
288 :
289 : // In an AddRec, check if both start and step are divisible.
290 4664 : if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
291 87 : const SCEV *Step = A->getStepRecurrence(SE);
292 87 : const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
293 87 : if (!FactorOutConstant(Step, StepRem, Factor, SE, DL))
294 : return false;
295 61 : if (!StepRem->isZero())
296 : return false;
297 61 : const SCEV *Start = A->getStart();
298 61 : if (!FactorOutConstant(Start, Remainder, Factor, SE, DL))
299 : return false;
300 122 : S = SE.getAddRecExpr(Start, Step, A->getLoop(),
301 : A->getNoWrapFlags(SCEV::FlagNW));
302 61 : return true;
303 : }
304 :
305 : return false;
306 : }
307 :
308 : /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
309 : /// is the number of SCEVAddRecExprs present, which are kept at the end of
310 : /// the list.
311 : ///
312 13282 : static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
313 : Type *Ty,
314 : ScalarEvolution &SE) {
315 : unsigned NumAddRecs = 0;
316 13350 : for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
317 68 : ++NumAddRecs;
318 : // Group Ops into non-addrecs and addrecs.
319 13282 : SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
320 13282 : SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
321 : // Let ScalarEvolution sort and simplify the non-addrecs list.
322 13282 : const SCEV *Sum = NoAddRecs.empty() ?
323 12822 : SE.getConstant(Ty, 0) :
324 13282 : SE.getAddExpr(NoAddRecs);
325 : // If it returned an add, use the operands. Otherwise it simplified
326 : // the sum into a single value, so just use that.
327 : Ops.clear();
328 : if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
329 14 : Ops.append(Add->op_begin(), Add->op_end());
330 13275 : else if (!Sum->isZero())
331 453 : Ops.push_back(Sum);
332 : // Then append the addrecs.
333 13282 : Ops.append(AddRecs.begin(), AddRecs.end());
334 13282 : }
335 :
336 : /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
337 : /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
338 : /// This helps expose more opportunities for folding parts of the expressions
339 : /// into GEP indices.
340 : ///
341 15421 : static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
342 : Type *Ty,
343 : ScalarEvolution &SE) {
344 : // Find the addrecs.
345 : SmallVector<const SCEV *, 8> AddRecs;
346 31924 : for (unsigned i = 0, e = Ops.size(); i != e; ++i)
347 33124 : while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
348 69 : const SCEV *Start = A->getStart();
349 69 : if (Start->isZero()) break;
350 59 : const SCEV *Zero = SE.getConstant(Ty, 0);
351 118 : AddRecs.push_back(SE.getAddRecExpr(Zero,
352 : A->getStepRecurrence(SE),
353 : A->getLoop(),
354 : A->getNoWrapFlags(SCEV::FlagNW)));
355 : if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
356 4 : Ops[i] = Zero;
357 8 : Ops.append(Add->op_begin(), Add->op_end());
358 4 : e += Add->getNumOperands();
359 : } else {
360 55 : Ops[i] = Start;
361 : }
362 : }
363 15421 : if (!AddRecs.empty()) {
364 : // Add the addrecs onto the end of the list.
365 59 : Ops.append(AddRecs.begin(), AddRecs.end());
366 : // Resort the operand list, moving any constants to the front.
367 59 : SimplifyAddOperands(Ops, Ty, SE);
368 : }
369 15421 : }
370 :
371 : /// expandAddToGEP - Expand an addition expression with a pointer type into
372 : /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
373 : /// BasicAliasAnalysis and other passes analyze the result. See the rules
374 : /// for getelementptr vs. inttoptr in
375 : /// http://llvm.org/docs/LangRef.html#pointeraliasing
376 : /// for details.
377 : ///
378 : /// Design note: The correctness of using getelementptr here depends on
379 : /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
380 : /// they may introduce pointer arithmetic which may not be safely converted
381 : /// into getelementptr.
382 : ///
383 : /// Design note: It might seem desirable for this function to be more
384 : /// loop-aware. If some of the indices are loop-invariant while others
385 : /// aren't, it might seem desirable to emit multiple GEPs, keeping the
386 : /// loop-invariant portions of the overall computation outside the loop.
387 : /// However, there are a few reasons this is not done here. Hoisting simple
388 : /// arithmetic is a low-level optimization that often isn't very
389 : /// important until late in the optimization process. In fact, passes
390 : /// like InstructionCombining will combine GEPs, even if it means
391 : /// pushing loop-invariant computation down into loops, so even if the
392 : /// GEPs were split here, the work would quickly be undone. The
393 : /// LoopStrengthReduction pass, which is usually run quite late (and
394 : /// after the last InstructionCombining pass), takes care of hoisting
395 : /// loop-invariant portions of expressions, after considering what
396 : /// can be folded using target addressing modes.
397 : ///
398 15421 : Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
399 : const SCEV *const *op_end,
400 : PointerType *PTy,
401 : Type *Ty,
402 : Value *V) {
403 15421 : Type *OriginalElTy = PTy->getElementType();
404 : Type *ElTy = OriginalElTy;
405 : SmallVector<Value *, 4> GepIndices;
406 : SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
407 : bool AnyNonZeroIndices = false;
408 :
409 : // Split AddRecs up into parts as either of the parts may be usable
410 : // without the other.
411 15421 : SplitAddRecs(Ops, Ty, SE);
412 :
413 15421 : Type *IntPtrTy = DL.getIntPtrType(PTy);
414 :
415 : // Descend down the pointer's type and attempt to convert the other
416 : // operands into GEP indices, at each level. The first index in a GEP
417 : // indexes into the array implied by the pointer operand; the rest of
418 : // the indices index into the element or field type selected by the
419 : // preceding index.
420 : for (;;) {
421 : // If the scale size is not 0, attempt to factor out a scale for
422 : // array indexing.
423 : SmallVector<const SCEV *, 8> ScaledOps;
424 17379 : if (ElTy->isSized()) {
425 17379 : const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy);
426 17379 : if (!ElSize->isZero()) {
427 : SmallVector<const SCEV *, 8> NewOps;
428 35934 : for (const SCEV *Op : Ops) {
429 18575 : const SCEV *Remainder = SE.getConstant(Ty, 0);
430 18575 : if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) {
431 : // Op now has ElSize factored out.
432 13998 : ScaledOps.push_back(Op);
433 13998 : if (!Remainder->isZero())
434 88 : NewOps.push_back(Remainder);
435 : AnyNonZeroIndices = true;
436 : } else {
437 : // The operand was not divisible, so add it to the list of operands
438 : // we'll scan next iteration.
439 4577 : NewOps.push_back(Op);
440 : }
441 : }
442 : // If we made any changes, update Ops.
443 17359 : if (!ScaledOps.empty()) {
444 : Ops = NewOps;
445 13223 : SimplifyAddOperands(Ops, Ty, SE);
446 : }
447 : }
448 : }
449 :
450 : // Record the scaled array index for this level of the type. If
451 : // we didn't find any operands that could be factored, tentatively
452 : // assume that element zero was selected (since the zero offset
453 : // would obviously be folded away).
454 17379 : Value *Scaled = ScaledOps.empty() ?
455 4156 : Constant::getNullValue(Ty) :
456 13223 : expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
457 17379 : GepIndices.push_back(Scaled);
458 :
459 : // Collect struct field index operands.
460 : while (StructType *STy = dyn_cast<StructType>(ElTy)) {
461 : bool FoundFieldNo = false;
462 : // An empty struct has no fields.
463 2253 : if (STy->getNumElements() == 0) break;
464 : // Field offsets are known. See if a constant offset falls within any of
465 : // the struct fields.
466 2251 : if (Ops.empty())
467 : break;
468 1769 : if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
469 1570 : if (SE.getTypeSizeInBits(C->getType()) <= 64) {
470 785 : const StructLayout &SL = *DL.getStructLayout(STy);
471 785 : uint64_t FullOffset = C->getValue()->getZExtValue();
472 785 : if (FullOffset < SL.getSizeInBytes()) {
473 678 : unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
474 678 : GepIndices.push_back(
475 678 : ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
476 678 : ElTy = STy->getTypeAtIndex(ElIdx);
477 678 : Ops[0] =
478 2034 : SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
479 : AnyNonZeroIndices = true;
480 : FoundFieldNo = true;
481 : }
482 : }
483 : // If no struct field offsets were found, tentatively assume that
484 : // field zero was selected (since the zero offset would obviously
485 : // be folded away).
486 1769 : if (!FoundFieldNo) {
487 1091 : ElTy = STy->getTypeAtIndex(0u);
488 1091 : GepIndices.push_back(
489 1091 : Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
490 : }
491 : }
492 :
493 : if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
494 1958 : ElTy = ATy->getElementType();
495 : else
496 : break;
497 : }
498 :
499 : // If none of the operands were convertible to proper GEP indices, cast
500 : // the base to i8* and do an ugly getelementptr with that. It's still
501 : // better than ptrtoint+arithmetic+inttoptr at least.
502 15421 : if (!AnyNonZeroIndices) {
503 : // Cast the base to i8*.
504 2049 : V = InsertNoopCastOfTo(V,
505 2049 : Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace()));
506 :
507 : assert(!isa<Instruction>(V) ||
508 : SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint()));
509 :
510 : // Expand the operands for a plain byte offset.
511 2049 : Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
512 :
513 : // Fold a GEP with constant operands.
514 : if (Constant *CLHS = dyn_cast<Constant>(V))
515 : if (Constant *CRHS = dyn_cast<Constant>(Idx))
516 12 : return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()),
517 : CLHS, CRHS);
518 :
519 : // Do a quick scan to see if we have this GEP nearby. If so, reuse it.
520 : unsigned ScanLimit = 6;
521 2037 : BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
522 : // Scanning starts from the last instruction before the insertion point.
523 2037 : BasicBlock::iterator IP = Builder.GetInsertPoint();
524 2037 : if (IP != BlockBegin) {
525 : --IP;
526 12059 : for (; ScanLimit; --IP, --ScanLimit) {
527 : // Don't count dbg.value against the ScanLimit, to avoid perturbing the
528 : // generated code.
529 11166 : if (isa<DbgInfoIntrinsic>(IP))
530 1730 : ScanLimit++;
531 2988 : if (IP->getOpcode() == Instruction::GetElementPtr &&
532 11395 : IP->getOperand(0) == V && IP->getOperand(1) == Idx)
533 : return &*IP;
534 11055 : if (IP == BlockBegin) break;
535 : }
536 : }
537 :
538 : // Save the original insertion point so we can restore it when we're done.
539 3852 : SCEVInsertPointGuard Guard(Builder, this);
540 :
541 : // Move the insertion point out of as many loops as we can.
542 3464 : while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
543 1526 : if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
544 12 : BasicBlock *Preheader = L->getLoopPreheader();
545 12 : if (!Preheader) break;
546 :
547 : // Ok, move up a level.
548 12 : Builder.SetInsertPoint(Preheader->getTerminator());
549 12 : }
550 :
551 : // Emit a GEP.
552 3852 : Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep");
553 1926 : rememberInstruction(GEP);
554 :
555 : return GEP;
556 : }
557 :
558 : {
559 26744 : SCEVInsertPointGuard Guard(Builder, this);
560 :
561 : // Move the insertion point out of as many loops as we can.
562 24683 : while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) {
563 10984 : if (!L->isLoopInvariant(V)) break;
564 :
565 : bool AnyIndexNotLoopInvariant = any_of(
566 0 : GepIndices, [L](Value *Op) { return !L->isLoopInvariant(Op); });
567 :
568 4112 : if (AnyIndexNotLoopInvariant)
569 : break;
570 :
571 327 : BasicBlock *Preheader = L->getLoopPreheader();
572 327 : if (!Preheader) break;
573 :
574 : // Ok, move up a level.
575 327 : Builder.SetInsertPoint(Preheader->getTerminator());
576 327 : }
577 :
578 : // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
579 : // because ScalarEvolution may have changed the address arithmetic to
580 : // compute a value which is beyond the end of the allocated object.
581 : Value *Casted = V;
582 13372 : if (V->getType() != PTy)
583 104 : Casted = InsertNoopCastOfTo(Casted, PTy);
584 26744 : Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep");
585 13372 : Ops.push_back(SE.getUnknown(GEP));
586 13372 : rememberInstruction(GEP);
587 : }
588 :
589 13372 : return expand(SE.getAddExpr(Ops));
590 : }
591 :
592 1061 : Value *SCEVExpander::expandAddToGEP(const SCEV *Op, PointerType *PTy, Type *Ty,
593 : Value *V) {
594 1061 : const SCEV *const Ops[1] = {Op};
595 1061 : return expandAddToGEP(Ops, Ops + 1, PTy, Ty, V);
596 : }
597 :
598 : /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
599 : /// SCEV expansion. If they are nested, this is the most nested. If they are
600 : /// neighboring, pick the later.
601 18871 : static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
602 : DominatorTree &DT) {
603 18871 : if (!A) return B;
604 1106 : if (!B) return A;
605 890 : if (A->contains(B)) return B;
606 684 : if (B->contains(A)) return A;
607 540 : if (DT.dominates(A->getHeader(), B->getHeader())) return B;
608 159 : if (DT.dominates(B->getHeader(), A->getHeader())) return A;
609 : return A; // Arbitrarily break the tie.
610 : }
611 :
612 : /// getRelevantLoop - Get the most relevant loop associated with the given
613 : /// expression, according to PickMostRelevantLoop.
614 61137 : const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) {
615 : // Test whether we've already computed the most relevant loop for this SCEV.
616 61137 : auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr));
617 61137 : if (!Pair.second)
618 22427 : return Pair.first->second;
619 :
620 38710 : if (isa<SCEVConstant>(S))
621 : // A constant has no relevant loops.
622 : return nullptr;
623 20696 : if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
624 : if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
625 23996 : return Pair.first->second = SE.LI.getLoopFor(I->getParent());
626 : // A non-instruction has no relevant loops.
627 : return nullptr;
628 : }
629 : if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
630 : const Loop *L = nullptr;
631 : if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
632 145 : L = AR->getLoop();
633 21947 : for (const SCEV *Op : N->operands())
634 14974 : L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT);
635 6973 : return RelevantLoops[N] = L;
636 : }
637 : if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) {
638 1123 : const Loop *Result = getRelevantLoop(C->getOperand());
639 1123 : return RelevantLoops[C] = Result;
640 : }
641 : if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
642 1184 : const Loop *Result = PickMostRelevantLoop(
643 1184 : getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT);
644 1184 : return RelevantLoops[D] = Result;
645 : }
646 0 : llvm_unreachable("Unexpected SCEV type!");
647 : }
648 :
649 : namespace {
650 :
651 : /// LoopCompare - Compare loops by PickMostRelevantLoop.
652 : class LoopCompare {
653 : DominatorTree &DT;
654 : public:
655 : explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
656 :
657 0 : bool operator()(std::pair<const Loop *, const SCEV *> LHS,
658 : std::pair<const Loop *, const SCEV *> RHS) const {
659 : // Keep pointer operands sorted at the end.
660 0 : if (LHS.second->getType()->isPointerTy() !=
661 0 : RHS.second->getType()->isPointerTy())
662 0 : return LHS.second->getType()->isPointerTy();
663 :
664 : // Compare loops with PickMostRelevantLoop.
665 0 : if (LHS.first != RHS.first)
666 0 : return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
667 :
668 : // If one operand is a non-constant negative and the other is not,
669 : // put the non-constant negative on the right so that a sub can
670 : // be used instead of a negate and add.
671 0 : if (LHS.second->isNonConstantNegative()) {
672 0 : if (!RHS.second->isNonConstantNegative())
673 0 : return false;
674 0 : } else if (RHS.second->isNonConstantNegative())
675 0 : return true;
676 :
677 : // Otherwise they are equivalent according to this comparison.
678 : return false;
679 : }
680 : };
681 :
682 : }
683 :
684 18355 : Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
685 18355 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
686 :
687 : // Collect all the add operands in a loop, along with their associated loops.
688 : // Iterate in reverse so that constants are emitted last, all else equal, and
689 : // so that pointer operands are inserted first, which the code below relies on
690 : // to form more involved GEPs.
691 : SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
692 18355 : for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
693 56864 : E(S->op_begin()); I != E; ++I)
694 38509 : OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
695 :
696 : // Sort by loop. Use a stable sort so that constants follow non-constants and
697 : // pointer operands precede non-pointer operands.
698 18355 : std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
699 :
700 : // Emit instructions to add all the operands. Hoist as much as possible
701 : // out of loops, and form meaningful getelementptrs where possible.
702 : Value *Sum = nullptr;
703 55792 : for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) {
704 37437 : const Loop *CurLoop = I->first;
705 37437 : const SCEV *Op = I->second;
706 37437 : if (!Sum) {
707 : // This is the first operand. Just expand it.
708 18355 : Sum = expand(Op);
709 18355 : ++I;
710 19082 : } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
711 : // The running sum expression is a pointer. Try to form a getelementptr
712 : // at this level with that as the base.
713 : SmallVector<const SCEV *, 4> NewOps;
714 29792 : for (; I != E && I->first == CurLoop; ++I) {
715 : // If the operand is SCEVUnknown and not instructions, peek through
716 : // it, to enable more of it to be folded into the GEP.
717 15432 : const SCEV *X = I->second;
718 8625 : if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
719 8625 : if (!isa<Instruction>(U->getValue()))
720 5605 : X = SE.getSCEV(U->getValue());
721 15432 : NewOps.push_back(X);
722 : }
723 14360 : Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
724 4722 : } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
725 : // The running sum is an integer, and there's a pointer at this level.
726 : // Try to form a getelementptr. If the running sum is instructions,
727 : // use a SCEVUnknown to avoid re-analyzing them.
728 : SmallVector<const SCEV *, 4> NewOps;
729 0 : NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
730 0 : SE.getSCEV(Sum));
731 0 : for (++I; I != E && I->first == CurLoop; ++I)
732 0 : NewOps.push_back(I->second);
733 0 : Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
734 4722 : } else if (Op->isNonConstantNegative()) {
735 : // Instead of doing a negate and add, just do a subtract.
736 998 : Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
737 998 : Sum = InsertNoopCastOfTo(Sum, Ty);
738 998 : Sum = InsertBinop(Instruction::Sub, Sum, W);
739 998 : ++I;
740 : } else {
741 : // A simple add.
742 3724 : Value *W = expandCodeFor(Op, Ty);
743 3724 : Sum = InsertNoopCastOfTo(Sum, Ty);
744 : // Canonicalize a constant to the RHS.
745 3724 : if (isa<Constant>(Sum)) std::swap(Sum, W);
746 3724 : Sum = InsertBinop(Instruction::Add, Sum, W);
747 3724 : ++I;
748 : }
749 : }
750 :
751 18355 : return Sum;
752 : }
753 :
754 1777 : Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
755 1777 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
756 :
757 : // Collect all the mul operands in a loop, along with their associated loops.
758 : // Iterate in reverse so that constants are emitted last, all else equal.
759 : SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
760 1777 : for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
761 5940 : E(S->op_begin()); I != E; ++I)
762 4163 : OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I));
763 :
764 : // Sort by loop. Use a stable sort so that constants follow non-constants.
765 1777 : std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT));
766 :
767 : // Emit instructions to mul all the operands. Hoist as much as possible
768 : // out of loops.
769 : Value *Prod = nullptr;
770 1777 : auto I = OpsAndLoops.begin();
771 :
772 : // Expand the calculation of X pow N in the following manner:
773 : // Let N = P1 + P2 + ... + PK, where all P are powers of 2. Then:
774 : // X pow N = (X pow P1) * (X pow P2) * ... * (X pow PK).
775 : const auto ExpandOpBinPowN = [this, &I, &OpsAndLoops, &Ty]() {
776 : auto E = I;
777 : // Calculate how many times the same operand from the same loop is included
778 : // into this power.
779 : uint64_t Exponent = 0;
780 : const uint64_t MaxExponent = UINT64_MAX >> 1;
781 : // No one sane will ever try to calculate such huge exponents, but if we
782 : // need this, we stop on UINT64_MAX / 2 because we need to exit the loop
783 : // below when the power of 2 exceeds our Exponent, and we want it to be
784 : // 1u << 31 at most to not deal with unsigned overflow.
785 : while (E != OpsAndLoops.end() && *I == *E && Exponent != MaxExponent) {
786 : ++Exponent;
787 : ++E;
788 : }
789 : assert(Exponent > 0 && "Trying to calculate a zeroth exponent of operand?");
790 :
791 : // Calculate powers with exponents 1, 2, 4, 8 etc. and include those of them
792 : // that are needed into the result.
793 : Value *P = expandCodeFor(I->second, Ty);
794 : Value *Result = nullptr;
795 : if (Exponent & 1)
796 : Result = P;
797 : for (uint64_t BinExp = 2; BinExp <= Exponent; BinExp <<= 1) {
798 : P = InsertBinop(Instruction::Mul, P, P);
799 : if (Exponent & BinExp)
800 : Result = Result ? InsertBinop(Instruction::Mul, Result, P) : P;
801 : }
802 :
803 : I = E;
804 : assert(Result && "Nothing was expanded?");
805 : return Result;
806 1777 : };
807 :
808 5433 : while (I != OpsAndLoops.end()) {
809 3656 : if (!Prod) {
810 : // This is the first operand. Just expand it.
811 1777 : Prod = ExpandOpBinPowN();
812 1879 : } else if (I->second->isAllOnesValue()) {
813 : // Instead of doing a multiply by negative one, just do a negate.
814 502 : Prod = InsertNoopCastOfTo(Prod, Ty);
815 502 : Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
816 502 : ++I;
817 : } else {
818 : // A simple mul.
819 1377 : Value *W = ExpandOpBinPowN();
820 1377 : Prod = InsertNoopCastOfTo(Prod, Ty);
821 : // Canonicalize a constant to the RHS.
822 1377 : if (isa<Constant>(Prod)) std::swap(Prod, W);
823 : const APInt *RHS;
824 1377 : if (match(W, m_Power2(RHS))) {
825 : // Canonicalize Prod*(1<<C) to Prod<<C.
826 : assert(!Ty->isVectorTy() && "vector types are not SCEVable");
827 694 : Prod = InsertBinop(Instruction::Shl, Prod,
828 1388 : ConstantInt::get(Ty, RHS->logBase2()));
829 : } else {
830 683 : Prod = InsertBinop(Instruction::Mul, Prod, W);
831 : }
832 : }
833 : }
834 :
835 1777 : return Prod;
836 : }
837 :
838 531 : Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
839 531 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
840 :
841 531 : Value *LHS = expandCodeFor(S->getLHS(), Ty);
842 531 : if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
843 : const APInt &RHS = SC->getAPInt();
844 510 : if (RHS.isPowerOf2())
845 388 : return InsertBinop(Instruction::LShr, LHS,
846 776 : ConstantInt::get(Ty, RHS.logBase2()));
847 : }
848 :
849 143 : Value *RHS = expandCodeFor(S->getRHS(), Ty);
850 143 : return InsertBinop(Instruction::UDiv, LHS, RHS);
851 : }
852 :
853 : /// Move parts of Base into Rest to leave Base with the minimal
854 : /// expression that provides a pointer operand suitable for a
855 : /// GEP expansion.
856 232 : static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
857 : ScalarEvolution &SE) {
858 233 : while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
859 1 : Base = A->getStart();
860 3 : Rest = SE.getAddExpr(Rest,
861 : SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
862 : A->getStepRecurrence(SE),
863 : A->getLoop(),
864 : A->getNoWrapFlags(SCEV::FlagNW)));
865 1 : }
866 : if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
867 118 : Base = A->getOperand(A->getNumOperands()-1);
868 : SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
869 59 : NewAddOps.back() = Rest;
870 59 : Rest = SE.getAddExpr(NewAddOps);
871 59 : ExposePointerBase(Base, Rest, SE);
872 : }
873 232 : }
874 :
875 : /// Determine if this is a well-behaved chain of instructions leading back to
876 : /// the PHI. If so, it may be reused by expanded expressions.
877 12 : bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV,
878 : const Loop *L) {
879 16 : if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) ||
880 1 : (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV)))
881 : return false;
882 : // If any of the operands don't dominate the insert position, bail.
883 : // Addrec operands are always loop-invariant, so this can only happen
884 : // if there are instructions which haven't been hoisted.
885 15 : if (L == IVIncInsertLoop) {
886 0 : for (User::op_iterator OI = IncV->op_begin()+1,
887 0 : OE = IncV->op_end(); OI != OE; ++OI)
888 : if (Instruction *OInst = dyn_cast<Instruction>(OI))
889 0 : if (!SE.DT.dominates(OInst, IVIncInsertPos))
890 : return false;
891 : }
892 : // Advance to the next instruction.
893 15 : IncV = dyn_cast<Instruction>(IncV->getOperand(0));
894 : if (!IncV)
895 : return false;
896 :
897 13 : if (IncV->mayHaveSideEffects())
898 : return false;
899 :
900 13 : if (IncV == PN)
901 : return true;
902 :
903 : return isNormalAddRecExprPHI(PN, IncV, L);
904 : }
905 :
906 : /// getIVIncOperand returns an induction variable increment's induction
907 : /// variable operand.
908 : ///
909 : /// If allowScale is set, any type of GEP is allowed as long as the nonIV
910 : /// operands dominate InsertPos.
911 : ///
912 : /// If allowScale is not set, ensure that a GEP increment conforms to one of the
913 : /// simple patterns generated by getAddRecExprPHILiterally and
914 : /// expandAddtoGEP. If the pattern isn't recognized, return NULL.
915 6284 : Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV,
916 : Instruction *InsertPos,
917 : bool allowScale) {
918 6284 : if (IncV == InsertPos)
919 : return nullptr;
920 :
921 6280 : switch (IncV->getOpcode()) {
922 : default:
923 : return nullptr;
924 : // Check for a simple Add/Sub or GEP of a loop invariant step.
925 4339 : case Instruction::Add:
926 : case Instruction::Sub: {
927 4339 : Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1));
928 128 : if (!OInst || SE.DT.dominates(OInst, InsertPos))
929 : return dyn_cast<Instruction>(IncV->getOperand(0));
930 : return nullptr;
931 : }
932 57 : case Instruction::BitCast:
933 57 : return dyn_cast<Instruction>(IncV->getOperand(0));
934 1857 : case Instruction::GetElementPtr:
935 5545 : for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) {
936 1902 : if (isa<Constant>(*I))
937 : continue;
938 : if (Instruction *OInst = dyn_cast<Instruction>(*I)) {
939 52 : if (!SE.DT.dominates(OInst, InsertPos))
940 : return nullptr;
941 : }
942 71 : if (allowScale) {
943 : // allow any kind of GEP as long as it can be hoisted.
944 : continue;
945 : }
946 : // This must be a pointer addition of constants (pretty), which is already
947 : // handled, or some number of address-size elements (ugly). Ugly geps
948 : // have 2 operands. i1* is used by the expander to represent an
949 : // address-size element.
950 70 : if (IncV->getNumOperands() != 2)
951 : return nullptr;
952 70 : unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace();
953 70 : if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS)
954 70 : && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS))
955 : return nullptr;
956 : break;
957 1786 : }
958 : return dyn_cast<Instruction>(IncV->getOperand(0));
959 : }
960 : }
961 :
962 : /// If the insert point of the current builder or any of the builders on the
963 : /// stack of saved builders has 'I' as its insert point, update it to point to
964 : /// the instruction after 'I'. This is intended to be used when the instruction
965 : /// 'I' is being moved. If this fixup is not done and 'I' is moved to a
966 : /// different block, the inconsistent insert point (with a mismatched
967 : /// Instruction and Block) can lead to an instruction being inserted in a block
968 : /// other than its parent.
969 288 : void SCEVExpander::fixupInsertPoints(Instruction *I) {
970 : BasicBlock::iterator It(*I);
971 : BasicBlock::iterator NewInsertPt = std::next(It);
972 288 : if (Builder.GetInsertPoint() == It)
973 2 : Builder.SetInsertPoint(&*NewInsertPt);
974 289 : for (auto *InsertPtGuard : InsertPointGuards)
975 1 : if (InsertPtGuard->GetInsertPoint() == It)
976 : InsertPtGuard->SetInsertPoint(NewInsertPt);
977 288 : }
978 :
979 : /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make
980 : /// it available to other uses in this loop. Recursively hoist any operands,
981 : /// until we reach a value that dominates InsertPos.
982 6087 : bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) {
983 6087 : if (SE.DT.dominates(IncV, InsertPos))
984 : return true;
985 :
986 : // InsertPos must itself dominate IncV so that IncV's new position satisfies
987 : // its existing users.
988 589 : if (isa<PHINode>(InsertPos) ||
989 291 : !SE.DT.dominates(InsertPos->getParent(), IncV->getParent()))
990 7 : return false;
991 :
992 291 : if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos))
993 : return false;
994 :
995 : // Check that the chain of IV operands leading back to Phi can be hoisted.
996 : SmallVector<Instruction*, 4> IVIncs;
997 : for(;;) {
998 296 : Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true);
999 296 : if (!Oper)
1000 : return false;
1001 : // IncV is safe to hoist.
1002 292 : IVIncs.push_back(IncV);
1003 292 : IncV = Oper;
1004 292 : if (SE.DT.dominates(IncV, InsertPos))
1005 : break;
1006 : }
1007 575 : for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) {
1008 288 : fixupInsertPoints(*I);
1009 288 : (*I)->moveBefore(InsertPos);
1010 : }
1011 : return true;
1012 : }
1013 :
1014 : /// Determine if this cyclic phi is in a form that would have been generated by
1015 : /// LSR. We don't care if the phi was actually expanded in this pass, as long
1016 : /// as it is in a low-cost form, for example, no implied multiplication. This
1017 : /// should match any patterns generated by getAddRecExprPHILiterally and
1018 : /// expandAddtoGEP.
1019 5914 : bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV,
1020 : const Loop *L) {
1021 : for(Instruction *IVOper = IncV;
1022 5988 : (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(),
1023 5988 : /*allowScale=*/false));) {
1024 5931 : if (IVOper == PN)
1025 : return true;
1026 : }
1027 : return false;
1028 : }
1029 :
1030 : /// expandIVInc - Expand an IV increment at Builder's current InsertPos.
1031 : /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may
1032 : /// need to materialize IV increments elsewhere to handle difficult situations.
1033 3647 : Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L,
1034 : Type *ExpandTy, Type *IntTy,
1035 : bool useSubtract) {
1036 : Value *IncV;
1037 : // If the PHI is a pointer, use a GEP, otherwise use an add or sub.
1038 3647 : if (ExpandTy->isPointerTy()) {
1039 : PointerType *GEPPtrTy = cast<PointerType>(ExpandTy);
1040 : // If the step isn't constant, don't use an implicitly scaled GEP, because
1041 : // that would require a multiply inside the loop.
1042 991 : if (!isa<ConstantInt>(StepV))
1043 104 : GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()),
1044 : GEPPtrTy->getAddressSpace());
1045 991 : IncV = expandAddToGEP(SE.getSCEV(StepV), GEPPtrTy, IntTy, PN);
1046 991 : if (IncV->getType() != PN->getType()) {
1047 434 : IncV = Builder.CreateBitCast(IncV, PN->getType());
1048 434 : rememberInstruction(IncV);
1049 : }
1050 : } else {
1051 2656 : IncV = useSubtract ?
1052 39 : Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") :
1053 7929 : Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next");
1054 2656 : rememberInstruction(IncV);
1055 : }
1056 3647 : return IncV;
1057 : }
1058 :
1059 : /// Hoist the addrec instruction chain rooted in the loop phi above the
1060 : /// position. This routine assumes that this is possible (has been checked).
1061 5742 : void SCEVExpander::hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist,
1062 : Instruction *Pos, PHINode *LoopPhi) {
1063 : do {
1064 5742 : if (DT->dominates(InstToHoist, Pos))
1065 : break;
1066 : // Make sure the increment is where we want it. But don't move it
1067 : // down past a potential existing post-inc user.
1068 0 : fixupInsertPoints(InstToHoist);
1069 0 : InstToHoist->moveBefore(Pos);
1070 : Pos = InstToHoist;
1071 0 : InstToHoist = cast<Instruction>(InstToHoist->getOperand(0));
1072 0 : } while (InstToHoist != LoopPhi);
1073 5742 : }
1074 :
1075 : /// Check whether we can cheaply express the requested SCEV in terms of
1076 : /// the available PHI SCEV by truncation and/or inversion of the step.
1077 22 : static bool canBeCheaplyTransformed(ScalarEvolution &SE,
1078 : const SCEVAddRecExpr *Phi,
1079 : const SCEVAddRecExpr *Requested,
1080 : bool &InvertStep) {
1081 22 : Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType());
1082 22 : Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType());
1083 :
1084 22 : if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth())
1085 : return false;
1086 :
1087 : // Try truncate it if necessary.
1088 22 : Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy));
1089 : if (!Phi)
1090 : return false;
1091 :
1092 : // Check whether truncation will help.
1093 22 : if (Phi == Requested) {
1094 0 : InvertStep = false;
1095 0 : return true;
1096 : }
1097 :
1098 : // Check whether inverting will help: {R,+,-1} == R - {0,+,1}.
1099 22 : if (SE.getAddExpr(Requested->getStart(),
1100 : SE.getNegativeSCEV(Requested)) == Phi) {
1101 2 : InvertStep = true;
1102 2 : return true;
1103 : }
1104 :
1105 : return false;
1106 : }
1107 :
1108 3633 : static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1109 3633 : if (!isa<IntegerType>(AR->getType()))
1110 : return false;
1111 :
1112 : unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1113 2641 : Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1114 2641 : const SCEV *Step = AR->getStepRecurrence(SE);
1115 2641 : const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy),
1116 : SE.getSignExtendExpr(AR, WideTy));
1117 : const SCEV *ExtendAfterOp =
1118 2641 : SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1119 2641 : return ExtendAfterOp == OpAfterExtend;
1120 : }
1121 :
1122 3633 : static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) {
1123 3633 : if (!isa<IntegerType>(AR->getType()))
1124 : return false;
1125 :
1126 : unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth();
1127 2641 : Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2);
1128 2641 : const SCEV *Step = AR->getStepRecurrence(SE);
1129 2641 : const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy),
1130 : SE.getZeroExtendExpr(AR, WideTy));
1131 : const SCEV *ExtendAfterOp =
1132 2641 : SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy);
1133 2641 : return ExtendAfterOp == OpAfterExtend;
1134 : }
1135 :
1136 : /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand
1137 : /// the base addrec, which is the addrec without any non-loop-dominating
1138 : /// values, and return the PHI.
1139 : PHINode *
1140 9461 : SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized,
1141 : const Loop *L,
1142 : Type *ExpandTy,
1143 : Type *IntTy,
1144 : Type *&TruncTy,
1145 : bool &InvertStep) {
1146 : assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position");
1147 :
1148 : // Reuse a previously-inserted PHI, if present.
1149 9461 : BasicBlock *LatchBlock = L->getLoopLatch();
1150 9461 : if (LatchBlock) {
1151 : PHINode *AddRecPhiMatch = nullptr;
1152 : Instruction *IncV = nullptr;
1153 9461 : TruncTy = nullptr;
1154 9461 : InvertStep = false;
1155 :
1156 : // Only try partially matching scevs that need truncation and/or
1157 : // step-inversion if we know this loop is outside the current loop.
1158 : bool TryNonMatchingSCEV =
1159 18628 : IVIncInsertLoop &&
1160 18334 : SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader());
1161 :
1162 29826 : for (PHINode &PN : L->getHeader()->phis()) {
1163 16717 : if (!SE.isSCEVable(PN.getType()))
1164 : continue;
1165 :
1166 16034 : const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&PN));
1167 : if (!PhiSCEV)
1168 : continue;
1169 :
1170 : bool IsMatchingSCEV = PhiSCEV == Normalized;
1171 : // We only handle truncation and inversion of phi recurrences for the
1172 : // expanded expression if the expanded expression's loop dominates the
1173 : // loop we insert to. Check now, so we can bail out early.
1174 13904 : if (!IsMatchingSCEV && !TryNonMatchingSCEV)
1175 : continue;
1176 :
1177 : // TODO: this possibly can be reworked to avoid this cast at all.
1178 : Instruction *TempIncV =
1179 5890 : dyn_cast<Instruction>(PN.getIncomingValueForBlock(LatchBlock));
1180 : if (!TempIncV)
1181 : continue;
1182 :
1183 : // Check whether we can reuse this PHI node.
1184 5889 : if (LSRMode) {
1185 5877 : if (!isExpandedAddRecExprPHI(&PN, TempIncV, L))
1186 : continue;
1187 5826 : if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos))
1188 : continue;
1189 : } else {
1190 12 : if (!isNormalAddRecExprPHI(&PN, TempIncV, L))
1191 : continue;
1192 : }
1193 :
1194 : // Stop if we have found an exact match SCEV.
1195 5835 : if (IsMatchingSCEV) {
1196 : IncV = TempIncV;
1197 5813 : TruncTy = nullptr;
1198 5813 : InvertStep = false;
1199 : AddRecPhiMatch = &PN;
1200 5813 : break;
1201 : }
1202 :
1203 : // Try whether the phi can be translated into the requested form
1204 : // (truncated and/or offset by a constant).
1205 44 : if ((!TruncTy || InvertStep) &&
1206 22 : canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) {
1207 : // Record the phi node. But don't stop we might find an exact match
1208 : // later.
1209 : AddRecPhiMatch = &PN;
1210 : IncV = TempIncV;
1211 2 : TruncTy = SE.getEffectiveSCEVType(Normalized->getType());
1212 : }
1213 : }
1214 :
1215 9461 : if (AddRecPhiMatch) {
1216 : // Potentially, move the increment. We have made sure in
1217 : // isExpandedAddRecExprPHI or hoistIVInc that this is possible.
1218 5815 : if (L == IVIncInsertLoop)
1219 5742 : hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch);
1220 :
1221 : // Ok, the add recurrence looks usable.
1222 : // Remember this PHI, even in post-inc mode.
1223 5815 : InsertedValues.insert(AddRecPhiMatch);
1224 : // Remember the increment.
1225 5815 : rememberInstruction(IncV);
1226 5815 : return AddRecPhiMatch;
1227 : }
1228 : }
1229 :
1230 : // Save the original insertion point so we can restore it when we're done.
1231 7292 : SCEVInsertPointGuard Guard(Builder, this);
1232 :
1233 : // Another AddRec may need to be recursively expanded below. For example, if
1234 : // this AddRec is quadratic, the StepV may itself be an AddRec in this
1235 : // loop. Remove this loop from the PostIncLoops set before expanding such
1236 : // AddRecs. Otherwise, we cannot find a valid position for the step
1237 : // (i.e. StepV can never dominate its loop header). Ideally, we could do
1238 : // SavedIncLoops.swap(PostIncLoops), but we generally have a single element,
1239 : // so it's not worth implementing SmallPtrSet::swap.
1240 : PostIncLoopSet SavedPostIncLoops = PostIncLoops;
1241 3646 : PostIncLoops.clear();
1242 :
1243 : // Expand code for the start value into the loop preheader.
1244 : assert(L->getLoopPreheader() &&
1245 : "Can't expand add recurrences without a loop preheader!");
1246 3646 : Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy,
1247 : L->getLoopPreheader()->getTerminator());
1248 :
1249 : // StartV must have been be inserted into L's preheader to dominate the new
1250 : // phi.
1251 : assert(!isa<Instruction>(StartV) ||
1252 : SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(),
1253 : L->getHeader()));
1254 :
1255 : // Expand code for the step value. Do this before creating the PHI so that PHI
1256 : // reuse code doesn't see an incomplete PHI.
1257 3646 : const SCEV *Step = Normalized->getStepRecurrence(SE);
1258 : // If the stride is negative, insert a sub instead of an add for the increment
1259 : // (unless it's a constant, because subtracts of constants are canonicalized
1260 : // to adds).
1261 3646 : bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1262 : if (useSubtract)
1263 13 : Step = SE.getNegativeSCEV(Step);
1264 : // Expand the step somewhere that dominates the loop header.
1265 3646 : Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1266 :
1267 : // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if
1268 : // we actually do emit an addition. It does not apply if we emit a
1269 : // subtraction.
1270 3646 : bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized);
1271 3646 : bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized);
1272 :
1273 : // Create the PHI.
1274 : BasicBlock *Header = L->getHeader();
1275 3646 : Builder.SetInsertPoint(Header, Header->begin());
1276 : pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1277 7292 : PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE),
1278 3646 : Twine(IVName) + ".iv");
1279 3646 : rememberInstruction(PN);
1280 :
1281 : // Create the step instructions and populate the PHI.
1282 10938 : for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1283 : BasicBlock *Pred = *HPI;
1284 :
1285 : // Add a start value.
1286 7292 : if (!L->contains(Pred)) {
1287 3646 : PN->addIncoming(StartV, Pred);
1288 : continue;
1289 : }
1290 :
1291 : // Create a step value and add it to the PHI.
1292 : // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the
1293 : // instructions at IVIncInsertPos.
1294 3646 : Instruction *InsertPos = L == IVIncInsertLoop ?
1295 : IVIncInsertPos : Pred->getTerminator();
1296 3646 : Builder.SetInsertPoint(InsertPos);
1297 3646 : Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1298 :
1299 3646 : if (isa<OverflowingBinaryOperator>(IncV)) {
1300 2655 : if (IncrementIsNUW)
1301 839 : cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap();
1302 2655 : if (IncrementIsNSW)
1303 1489 : cast<BinaryOperator>(IncV)->setHasNoSignedWrap();
1304 : }
1305 3646 : PN->addIncoming(IncV, Pred);
1306 : }
1307 :
1308 : // After expanding subexpressions, restore the PostIncLoops set so the caller
1309 : // can ensure that IVIncrement dominates the current uses.
1310 : PostIncLoops = SavedPostIncLoops;
1311 :
1312 : // Remember this PHI, even in post-inc mode.
1313 3646 : InsertedValues.insert(PN);
1314 :
1315 : return PN;
1316 : }
1317 :
1318 9461 : Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
1319 : Type *STy = S->getType();
1320 9461 : Type *IntTy = SE.getEffectiveSCEVType(STy);
1321 9461 : const Loop *L = S->getLoop();
1322 :
1323 : // Determine a normalized form of this expression, which is the expression
1324 : // before any post-inc adjustment is made.
1325 : const SCEVAddRecExpr *Normalized = S;
1326 9461 : if (PostIncLoops.count(L)) {
1327 : PostIncLoopSet Loops;
1328 3822 : Loops.insert(L);
1329 3822 : Normalized = cast<SCEVAddRecExpr>(normalizeForPostIncUse(S, Loops, SE));
1330 : }
1331 :
1332 : // Strip off any non-loop-dominating component from the addrec start.
1333 9461 : const SCEV *Start = Normalized->getStart();
1334 : const SCEV *PostLoopOffset = nullptr;
1335 18922 : if (!SE.properlyDominates(Start, L->getHeader())) {
1336 : PostLoopOffset = Start;
1337 1 : Start = SE.getConstant(Normalized->getType(), 0);
1338 1 : Normalized = cast<SCEVAddRecExpr>(
1339 1 : SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE),
1340 : Normalized->getLoop(),
1341 : Normalized->getNoWrapFlags(SCEV::FlagNW)));
1342 : }
1343 :
1344 : // Strip off any non-loop-dominating component from the addrec step.
1345 9461 : const SCEV *Step = Normalized->getStepRecurrence(SE);
1346 : const SCEV *PostLoopScale = nullptr;
1347 18922 : if (!SE.dominates(Step, L->getHeader())) {
1348 : PostLoopScale = Step;
1349 0 : Step = SE.getConstant(Normalized->getType(), 1);
1350 0 : if (!Start->isZero()) {
1351 : // The normalization below assumes that Start is constant zero, so if
1352 : // it isn't re-associate Start to PostLoopOffset.
1353 : assert(!PostLoopOffset && "Start not-null but PostLoopOffset set?");
1354 : PostLoopOffset = Start;
1355 0 : Start = SE.getConstant(Normalized->getType(), 0);
1356 : }
1357 : Normalized =
1358 0 : cast<SCEVAddRecExpr>(SE.getAddRecExpr(
1359 : Start, Step, Normalized->getLoop(),
1360 : Normalized->getNoWrapFlags(SCEV::FlagNW)));
1361 : }
1362 :
1363 : // Expand the core addrec. If we need post-loop scaling, force it to
1364 : // expand to an integer type to avoid the need for additional casting.
1365 9461 : Type *ExpandTy = PostLoopScale ? IntTy : STy;
1366 : // We can't use a pointer type for the addrec if the pointer type is
1367 : // non-integral.
1368 : Type *AddRecPHIExpandTy =
1369 9461 : DL.isNonIntegralPointerType(STy) ? Normalized->getType() : ExpandTy;
1370 :
1371 : // In some cases, we decide to reuse an existing phi node but need to truncate
1372 : // it and/or invert the step.
1373 9461 : Type *TruncTy = nullptr;
1374 9461 : bool InvertStep = false;
1375 9461 : PHINode *PN = getAddRecExprPHILiterally(Normalized, L, AddRecPHIExpandTy,
1376 : IntTy, TruncTy, InvertStep);
1377 :
1378 : // Accommodate post-inc mode, if necessary.
1379 : Value *Result;
1380 9461 : if (!PostIncLoops.count(L))
1381 : Result = PN;
1382 : else {
1383 : // In PostInc mode, use the post-incremented value.
1384 3822 : BasicBlock *LatchBlock = L->getLoopLatch();
1385 : assert(LatchBlock && "PostInc mode requires a unique loop latch!");
1386 3822 : Result = PN->getIncomingValueForBlock(LatchBlock);
1387 :
1388 : // For an expansion to use the postinc form, the client must call
1389 : // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
1390 : // or dominated by IVIncInsertPos.
1391 7644 : if (isa<Instruction>(Result) &&
1392 7644 : !SE.DT.dominates(cast<Instruction>(Result),
1393 : &*Builder.GetInsertPoint())) {
1394 : // The induction variable's postinc expansion does not dominate this use.
1395 : // IVUsers tries to prevent this case, so it is rare. However, it can
1396 : // happen when an IVUser outside the loop is not dominated by the latch
1397 : // block. Adjusting IVIncInsertPos before expansion begins cannot handle
1398 : // all cases. Consider a phi outside whose operand is replaced during
1399 : // expansion with the value of the postinc user. Without fundamentally
1400 : // changing the way postinc users are tracked, the only remedy is
1401 : // inserting an extra IV increment. StepV might fold into PostLoopOffset,
1402 : // but hopefully expandCodeFor handles that.
1403 : bool useSubtract =
1404 1 : !ExpandTy->isPointerTy() && Step->isNonConstantNegative();
1405 : if (useSubtract)
1406 0 : Step = SE.getNegativeSCEV(Step);
1407 : Value *StepV;
1408 : {
1409 : // Expand the step somewhere that dominates the loop header.
1410 2 : SCEVInsertPointGuard Guard(Builder, this);
1411 1 : StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front());
1412 : }
1413 1 : Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract);
1414 : }
1415 : }
1416 :
1417 : // We have decided to reuse an induction variable of a dominating loop. Apply
1418 : // truncation and/or inversion of the step.
1419 9461 : if (TruncTy) {
1420 2 : Type *ResTy = Result->getType();
1421 : // Normalize the result type.
1422 2 : if (ResTy != SE.getEffectiveSCEVType(ResTy))
1423 0 : Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy));
1424 : // Truncate the result.
1425 2 : if (TruncTy != Result->getType()) {
1426 2 : Result = Builder.CreateTrunc(Result, TruncTy);
1427 2 : rememberInstruction(Result);
1428 : }
1429 : // Invert the result.
1430 2 : if (InvertStep) {
1431 6 : Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy),
1432 : Result);
1433 2 : rememberInstruction(Result);
1434 : }
1435 : }
1436 :
1437 : // Re-apply any non-loop-dominating scale.
1438 9461 : if (PostLoopScale) {
1439 : assert(S->isAffine() && "Can't linearly scale non-affine recurrences.");
1440 0 : Result = InsertNoopCastOfTo(Result, IntTy);
1441 0 : Result = Builder.CreateMul(Result,
1442 : expandCodeFor(PostLoopScale, IntTy));
1443 0 : rememberInstruction(Result);
1444 : }
1445 :
1446 : // Re-apply any non-loop-dominating offset.
1447 9461 : if (PostLoopOffset) {
1448 : if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) {
1449 2 : if (Result->getType()->isIntegerTy()) {
1450 1 : Value *Base = expandCodeFor(PostLoopOffset, ExpandTy);
1451 1 : Result = expandAddToGEP(SE.getUnknown(Result), PTy, IntTy, Base);
1452 : } else {
1453 0 : Result = expandAddToGEP(PostLoopOffset, PTy, IntTy, Result);
1454 : }
1455 : } else {
1456 0 : Result = InsertNoopCastOfTo(Result, IntTy);
1457 0 : Result = Builder.CreateAdd(Result,
1458 : expandCodeFor(PostLoopOffset, IntTy));
1459 0 : rememberInstruction(Result);
1460 : }
1461 : }
1462 :
1463 9461 : return Result;
1464 : }
1465 :
1466 9843 : Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
1467 9843 : if (!CanonicalMode) return expandAddRecExprLiterally(S);
1468 :
1469 382 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
1470 382 : const Loop *L = S->getLoop();
1471 :
1472 : // First check for an existing canonical IV in a suitable type.
1473 : PHINode *CanonicalIV = nullptr;
1474 382 : if (PHINode *PN = L->getCanonicalInductionVariable())
1475 234 : if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
1476 : CanonicalIV = PN;
1477 :
1478 : // Rewrite an AddRec in terms of the canonical induction variable, if
1479 : // its type is more narrow.
1480 222 : if (CanonicalIV &&
1481 222 : SE.getTypeSizeInBits(CanonicalIV->getType()) >
1482 222 : SE.getTypeSizeInBits(Ty)) {
1483 11 : SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
1484 33 : for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
1485 44 : NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
1486 22 : Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(),
1487 : S->getNoWrapFlags(SCEV::FlagNW)));
1488 : BasicBlock::iterator NewInsertPt =
1489 11 : findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock());
1490 11 : V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr,
1491 11 : &*NewInsertPt);
1492 : return V;
1493 : }
1494 :
1495 : // {X,+,F} --> X + {0,+,F}
1496 742 : if (!S->getStart()->isZero()) {
1497 173 : SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
1498 346 : NewOps[0] = SE.getConstant(Ty, 0);
1499 346 : const SCEV *Rest = SE.getAddRecExpr(NewOps, L,
1500 : S->getNoWrapFlags(SCEV::FlagNW));
1501 :
1502 : // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
1503 : // comments on expandAddToGEP for details.
1504 173 : const SCEV *Base = S->getStart();
1505 : // Dig into the expression to find the pointer base for a GEP.
1506 173 : const SCEV *ExposedRest = Rest;
1507 173 : ExposePointerBase(Base, ExposedRest, SE);
1508 : // If we found a pointer, expand the AddRec with a GEP.
1509 173 : if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
1510 : // Make sure the Base isn't something exotic, such as a multiplied
1511 : // or divided pointer value. In those cases, the result type isn't
1512 : // actually a pointer type.
1513 130 : if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
1514 65 : Value *StartV = expand(Base);
1515 : assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
1516 65 : return expandAddToGEP(ExposedRest, PTy, Ty, StartV);
1517 : }
1518 : }
1519 :
1520 : // Just do a normal add. Pre-expand the operands to suppress folding.
1521 : //
1522 : // The LHS and RHS values are factored out of the expand call to make the
1523 : // output independent of the argument evaluation order.
1524 216 : const SCEV *AddExprLHS = SE.getUnknown(expand(S->getStart()));
1525 108 : const SCEV *AddExprRHS = SE.getUnknown(expand(Rest));
1526 108 : return expand(SE.getAddExpr(AddExprLHS, AddExprRHS));
1527 : }
1528 :
1529 : // If we don't yet have a canonical IV, create one.
1530 198 : if (!CanonicalIV) {
1531 : // Create and insert the PHI node for the induction variable in the
1532 : // specified loop.
1533 : BasicBlock *Header = L->getHeader();
1534 : pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header);
1535 87 : CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar",
1536 : &Header->front());
1537 87 : rememberInstruction(CanonicalIV);
1538 :
1539 : SmallSet<BasicBlock *, 4> PredSeen;
1540 87 : Constant *One = ConstantInt::get(Ty, 1);
1541 265 : for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
1542 : BasicBlock *HP = *HPI;
1543 178 : if (!PredSeen.insert(HP).second) {
1544 : // There must be an incoming value for each predecessor, even the
1545 : // duplicates!
1546 0 : CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
1547 0 : continue;
1548 : }
1549 :
1550 178 : if (L->contains(HP)) {
1551 : // Insert a unit add instruction right before the terminator
1552 : // corresponding to the back-edge.
1553 91 : Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
1554 : "indvar.next",
1555 : HP->getTerminator());
1556 91 : Add->setDebugLoc(HP->getTerminator()->getDebugLoc());
1557 91 : rememberInstruction(Add);
1558 91 : CanonicalIV->addIncoming(Add, HP);
1559 : } else {
1560 87 : CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
1561 : }
1562 : }
1563 : }
1564 :
1565 : // {0,+,1} --> Insert a canonical induction variable into the loop!
1566 198 : if (S->isAffine() && S->getOperand(1)->isOne()) {
1567 : assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
1568 : "IVs with types different from the canonical IV should "
1569 : "already have been handled!");
1570 : return CanonicalIV;
1571 : }
1572 :
1573 : // {0,+,F} --> {0,+,1} * F
1574 :
1575 : // If this is a simple linear addrec, emit it now as a special case.
1576 84 : if (S->isAffine()) // {0,+,F} --> i*F
1577 : return
1578 252 : expand(SE.getTruncateOrNoop(
1579 84 : SE.getMulExpr(SE.getUnknown(CanonicalIV),
1580 : SE.getNoopOrAnyExtend(S->getOperand(1),
1581 : CanonicalIV->getType())),
1582 84 : Ty));
1583 :
1584 : // If this is a chain of recurrences, turn it into a closed form, using the
1585 : // folders, then expandCodeFor the closed form. This allows the folders to
1586 : // simplify the expression without having to build a bunch of special code
1587 : // into this folder.
1588 0 : const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
1589 :
1590 : // Promote S up to the canonical IV type, if the cast is foldable.
1591 : const SCEV *NewS = S;
1592 0 : const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
1593 0 : if (isa<SCEVAddRecExpr>(Ext))
1594 : NewS = Ext;
1595 :
1596 0 : const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
1597 : //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
1598 :
1599 : // Truncate the result down to the original type, if needed.
1600 0 : const SCEV *T = SE.getTruncateOrNoop(V, Ty);
1601 0 : return expand(T);
1602 : }
1603 :
1604 218 : Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
1605 218 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
1606 218 : Value *V = expandCodeFor(S->getOperand(),
1607 218 : SE.getEffectiveSCEVType(S->getOperand()->getType()));
1608 218 : Value *I = Builder.CreateTrunc(V, Ty);
1609 218 : rememberInstruction(I);
1610 218 : return I;
1611 : }
1612 :
1613 414 : Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
1614 414 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
1615 414 : Value *V = expandCodeFor(S->getOperand(),
1616 414 : SE.getEffectiveSCEVType(S->getOperand()->getType()));
1617 414 : Value *I = Builder.CreateZExt(V, Ty);
1618 414 : rememberInstruction(I);
1619 414 : return I;
1620 : }
1621 :
1622 149 : Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
1623 149 : Type *Ty = SE.getEffectiveSCEVType(S->getType());
1624 149 : Value *V = expandCodeFor(S->getOperand(),
1625 149 : SE.getEffectiveSCEVType(S->getOperand()->getType()));
1626 149 : Value *I = Builder.CreateSExt(V, Ty);
1627 149 : rememberInstruction(I);
1628 149 : return I;
1629 : }
1630 :
1631 379 : Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
1632 758 : Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1633 379 : Type *Ty = LHS->getType();
1634 782 : for (int i = S->getNumOperands()-2; i >= 0; --i) {
1635 : // In the case of mixed integer and pointer types, do the
1636 : // rest of the comparisons as integer.
1637 806 : if (S->getOperand(i)->getType() != Ty) {
1638 0 : Ty = SE.getEffectiveSCEVType(Ty);
1639 0 : LHS = InsertNoopCastOfTo(LHS, Ty);
1640 : }
1641 806 : Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1642 403 : Value *ICmp = Builder.CreateICmpSGT(LHS, RHS);
1643 403 : rememberInstruction(ICmp);
1644 403 : Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
1645 403 : rememberInstruction(Sel);
1646 : LHS = Sel;
1647 : }
1648 : // In the case of mixed integer and pointer types, cast the
1649 : // final result back to the pointer type.
1650 379 : if (LHS->getType() != S->getType())
1651 0 : LHS = InsertNoopCastOfTo(LHS, S->getType());
1652 379 : return LHS;
1653 : }
1654 :
1655 108 : Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
1656 216 : Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
1657 108 : Type *Ty = LHS->getType();
1658 216 : for (int i = S->getNumOperands()-2; i >= 0; --i) {
1659 : // In the case of mixed integer and pointer types, do the
1660 : // rest of the comparisons as integer.
1661 216 : if (S->getOperand(i)->getType() != Ty) {
1662 3 : Ty = SE.getEffectiveSCEVType(Ty);
1663 3 : LHS = InsertNoopCastOfTo(LHS, Ty);
1664 : }
1665 216 : Value *RHS = expandCodeFor(S->getOperand(i), Ty);
1666 108 : Value *ICmp = Builder.CreateICmpUGT(LHS, RHS);
1667 108 : rememberInstruction(ICmp);
1668 108 : Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
1669 108 : rememberInstruction(Sel);
1670 : LHS = Sel;
1671 : }
1672 : // In the case of mixed integer and pointer types, cast the
1673 : // final result back to the pointer type.
1674 108 : if (LHS->getType() != S->getType())
1675 2 : LHS = InsertNoopCastOfTo(LHS, S->getType());
1676 108 : return LHS;
1677 : }
1678 :
1679 21284 : Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty,
1680 : Instruction *IP) {
1681 : setInsertPoint(IP);
1682 21284 : return expandCodeFor(SH, Ty);
1683 : }
1684 :
1685 105697 : Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) {
1686 : // Expand the code for this SCEV.
1687 105697 : Value *V = expand(SH);
1688 105697 : if (Ty) {
1689 : assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
1690 : "non-trivial casts should be done with the SCEVs directly!");
1691 78489 : V = InsertNoopCastOfTo(V, Ty);
1692 : }
1693 105697 : return V;
1694 : }
1695 :
1696 : ScalarEvolution::ValueOffsetPair
1697 100340 : SCEVExpander::FindValueInExprValueMap(const SCEV *S,
1698 : const Instruction *InsertPt) {
1699 100340 : SetVector<ScalarEvolution::ValueOffsetPair> *Set = SE.getSCEVValues(S);
1700 : // If the expansion is not in CanonicalMode, and the SCEV contains any
1701 : // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally.
1702 100340 : if (CanonicalMode || !SE.containsAddRecurrence(S)) {
1703 : // If S is scConstant, it may be worse to reuse an existing Value.
1704 90503 : if (S->getSCEVType() != scConstant && Set) {
1705 : // Choose a Value from the set which dominates the insertPt.
1706 : // insertPt should be inside the Value's parent loop so as not to break
1707 : // the LCSSA form.
1708 20921 : for (auto const &VOPair : *Set) {
1709 14134 : Value *V = VOPair.first;
1710 14134 : ConstantInt *Offset = VOPair.second;
1711 : Instruction *EntInst = nullptr;
1712 14134 : if (V && isa<Instruction>(V) && (EntInst = cast<Instruction>(V)) &&
1713 13843 : S->getType() == V->getType() &&
1714 12510 : EntInst->getFunction() == InsertPt->getFunction() &&
1715 21874 : SE.DT.dominates(EntInst, InsertPt) &&
1716 7448 : (SE.LI.getLoopFor(EntInst->getParent()) == nullptr ||
1717 1485 : SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt)))
1718 5741 : return {V, Offset};
1719 : }
1720 : }
1721 : }
1722 94599 : return {nullptr, nullptr};
1723 : }
1724 :
1725 : // The expansion of SCEV will either reuse a previous Value in ExprValueMap,
1726 : // or expand the SCEV literally. Specifically, if the expansion is in LSRMode,
1727 : // and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded
1728 : // literally, to prevent LSR's transformed SCEV from being reverted. Otherwise,
1729 : // the expansion will try to reuse Value from ExprValueMap, and only when it
1730 : // fails, expand the SCEV literally.
1731 138395 : Value *SCEVExpander::expand(const SCEV *S) {
1732 : // Compute an insertion point for this SCEV object. Hoist the instructions
1733 : // as far out in the loop nest as possible.
1734 : Instruction *InsertPt = &*Builder.GetInsertPoint();
1735 138395 : for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
1736 40159 : L = L->getParentLoop())
1737 178554 : if (SE.isLoopInvariant(S, L)) {
1738 81907 : if (!L) break;
1739 40159 : if (BasicBlock *Preheader = L->getLoopPreheader())
1740 : InsertPt = Preheader->getTerminator();
1741 : else {
1742 : // LSR sets the insertion point for AddRec start/step values to the
1743 : // block start to simplify value reuse, even though it's an invalid
1744 : // position. SCEVExpander must correct for this in all cases.
1745 : InsertPt = &*L->getHeader()->getFirstInsertionPt();
1746 : }
1747 : } else {
1748 : // We can move insertion point only if there is no div or rem operations
1749 : // otherwise we are risky to move it over the check for zero denominator.
1750 : auto SafeToHoist = [](const SCEV *S) {
1751 : return !SCEVExprContains(S, [](const SCEV *S) {
1752 : if (const auto *D = dyn_cast<SCEVUDivExpr>(S)) {
1753 : if (const auto *SC = dyn_cast<SCEVConstant>(D->getRHS()))
1754 : // Division by non-zero constants can be hoisted.
1755 : return SC->getValue()->isZero();
1756 : // All other divisions should not be moved as they may be
1757 : // divisions by zero and should be kept within the
1758 : // conditions of the surrounding loops that guard their
1759 : // execution (see PR35406).
1760 : return true;
1761 : }
1762 : return false;
1763 : });
1764 : };
1765 : // If the SCEV is computable at this level, insert it into the header
1766 : // after the PHIs (and after any other instructions that we've inserted
1767 : // there) so that it is guaranteed to dominate any user inside the loop.
1768 108782 : if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L) &&
1769 : SafeToHoist(S))
1770 : InsertPt = &*L->getHeader()->getFirstInsertionPt();
1771 154676 : while (InsertPt->getIterator() != Builder.GetInsertPoint() &&
1772 39038 : (isInsertedInstruction(InsertPt) ||
1773 : isa<DbgInfoIntrinsic>(InsertPt))) {
1774 : InsertPt = &*std::next(InsertPt->getIterator());
1775 : }
1776 : break;
1777 40159 : }
1778 :
1779 : // Check to see if we already expanded this here.
1780 138395 : auto I = InsertedExpressions.find(std::make_pair(S, InsertPt));
1781 138395 : if (I != InsertedExpressions.end())
1782 40637 : return I->second;
1783 :
1784 97758 : SCEVInsertPointGuard Guard(Builder, this);
1785 97758 : Builder.SetInsertPoint(InsertPt);
1786 :
1787 : // Expand the expression into instructions.
1788 97758 : ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, InsertPt);
1789 97758 : Value *V = VO.first;
1790 :
1791 97758 : if (!V)
1792 92335 : V = visit(S);
1793 5423 : else if (VO.second) {
1794 104 : if (PointerType *Vty = dyn_cast<PointerType>(V->getType())) {
1795 2 : Type *Ety = Vty->getPointerElementType();
1796 : int64_t Offset = VO.second->getSExtValue();
1797 2 : int64_t ESize = SE.getTypeSizeInBits(Ety);
1798 2 : if ((Offset * 8) % ESize == 0) {
1799 : ConstantInt *Idx =
1800 2 : ConstantInt::getSigned(VO.second->getType(), -(Offset * 8) / ESize);
1801 2 : V = Builder.CreateGEP(Ety, V, Idx, "scevgep");
1802 : } else {
1803 : ConstantInt *Idx =
1804 2 : ConstantInt::getSigned(VO.second->getType(), -Offset);
1805 : unsigned AS = Vty->getAddressSpace();
1806 3 : V = Builder.CreateBitCast(V, Type::getInt8PtrTy(SE.getContext(), AS));
1807 1 : V = Builder.CreateGEP(Type::getInt8Ty(SE.getContext()), V, Idx,
1808 : "uglygep");
1809 1 : V = Builder.CreateBitCast(V, Vty);
1810 : }
1811 : } else {
1812 204 : V = Builder.CreateSub(V, VO.second);
1813 : }
1814 : }
1815 : // Remember the expanded value for this SCEV at this location.
1816 : //
1817 : // This is independent of PostIncLoops. The mapped value simply materializes
1818 : // the expression at this insertion point. If the mapped value happened to be
1819 : // a postinc expansion, it could be reused by a non-postinc user, but only if
1820 : // its insertion point was already at the head of the loop.
1821 195516 : InsertedExpressions[std::make_pair(S, InsertPt)] = V;
1822 : return V;
1823 : }
1824 :
1825 46866 : void SCEVExpander::rememberInstruction(Value *I) {
1826 46866 : if (!PostIncLoops.empty())
1827 3921 : InsertedPostIncValues.insert(I);
1828 : else
1829 42945 : InsertedValues.insert(I);
1830 46866 : }
1831 :
1832 : /// getOrInsertCanonicalInductionVariable - This method returns the
1833 : /// canonical induction variable of the specified type for the specified
1834 : /// loop (inserting one if there is none). A canonical induction variable
1835 : /// starts at zero and steps by one on each iteration.
1836 : PHINode *
1837 0 : SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
1838 : Type *Ty) {
1839 : assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
1840 :
1841 : // Build a SCEV for {0,+,1}<L>.
1842 : // Conservatively use FlagAnyWrap for now.
1843 0 : const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
1844 : SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap);
1845 :
1846 : // Emit code for it.
1847 0 : SCEVInsertPointGuard Guard(Builder, this);
1848 : PHINode *V =
1849 0 : cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front()));
1850 :
1851 0 : return V;
1852 : }
1853 :
1854 : /// replaceCongruentIVs - Check for congruent phis in this loop header and
1855 : /// replace them with their most canonical representative. Return the number of
1856 : /// phis eliminated.
1857 : ///
1858 : /// This does not depend on any SCEVExpander state but should be used in
1859 : /// the same context that SCEVExpander is used.
1860 : unsigned
1861 13637 : SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT,
1862 : SmallVectorImpl<WeakTrackingVH> &DeadInsts,
1863 : const TargetTransformInfo *TTI) {
1864 : // Find integer phis in order of increasing width.
1865 : SmallVector<PHINode*, 8> Phis;
1866 49065 : for (PHINode &PN : L->getHeader()->phis())
1867 21791 : Phis.push_back(&PN);
1868 :
1869 13637 : if (TTI)
1870 : llvm::sort(Phis, [](Value *LHS, Value *RHS) {
1871 : // Put pointers at the back and make sure pointer < pointer = false.
1872 : if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy())
1873 : return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy();
1874 : return RHS->getType()->getPrimitiveSizeInBits() <
1875 : LHS->getType()->getPrimitiveSizeInBits();
1876 : });
1877 :
1878 : unsigned NumElim = 0;
1879 : DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1880 : // Process phis from wide to narrow. Map wide phis to their truncation
1881 : // so narrow phis can reuse them.
1882 35428 : for (PHINode *Phi : Phis) {
1883 : auto SimplifyPHINode = [&](PHINode *PN) -> Value * {
1884 : if (Value *V = SimplifyInstruction(PN, {DL, &SE.TLI, &SE.DT, &SE.AC}))
1885 : return V;
1886 : if (!SE.isSCEVable(PN->getType()))
1887 : return nullptr;
1888 : auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN));
1889 : if (!Const)
1890 : return nullptr;
1891 : return Const->getValue();
1892 21791 : };
1893 :
1894 : // Fold constant phis. They may be congruent to other constant phis and
1895 : // would confuse the logic below that expects proper IVs.
1896 21791 : if (Value *V = SimplifyPHINode(Phi)) {
1897 15 : if (V->getType() != Phi->getType())
1898 21732 : continue;
1899 14 : Phi->replaceAllUsesWith(V);
1900 14 : DeadInsts.emplace_back(Phi);
1901 14 : ++NumElim;
1902 : DEBUG_WITH_TYPE(DebugType, dbgs()
1903 : << "INDVARS: Eliminated constant iv: " << *Phi << '\n');
1904 14 : continue;
1905 : }
1906 :
1907 21776 : if (!SE.isSCEVable(Phi->getType()))
1908 : continue;
1909 :
1910 19601 : PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)];
1911 19601 : if (!OrigPhiRef) {
1912 19540 : OrigPhiRef = Phi;
1913 46474 : if (Phi->getType()->isIntegerTy() && TTI &&
1914 7394 : TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) {
1915 : // This phi can be freely truncated to the narrowest phi type. Map the
1916 : // truncated expression to it so it will be reused for narrow types.
1917 : const SCEV *TruncExpr =
1918 1280 : SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType());
1919 640 : ExprToIVMap[TruncExpr] = Phi;
1920 : }
1921 19540 : continue;
1922 : }
1923 :
1924 : // Replacing a pointer phi with an integer phi or vice-versa doesn't make
1925 : // sense.
1926 183 : if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy())
1927 : continue;
1928 :
1929 59 : if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1930 59 : Instruction *OrigInc = dyn_cast<Instruction>(
1931 : OrigPhiRef->getIncomingValueForBlock(LatchBlock));
1932 : Instruction *IsomorphicInc =
1933 59 : dyn_cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1934 :
1935 59 : if (OrigInc && IsomorphicInc) {
1936 : // If this phi has the same width but is more canonical, replace the
1937 : // original with it. As part of the "more canonical" determination,
1938 : // respect a prior decision to use an IV chain.
1939 57 : if (OrigPhiRef->getType() == Phi->getType() &&
1940 33 : !(ChainedPhis.count(Phi) ||
1941 90 : isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) &&
1942 61 : (ChainedPhis.count(Phi) ||
1943 4 : isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) {
1944 : std::swap(OrigPhiRef, Phi);
1945 : std::swap(OrigInc, IsomorphicInc);
1946 : }
1947 : // Replacing the congruent phi is sufficient because acyclic
1948 : // redundancy elimination, CSE/GVN, should handle the
1949 : // rest. However, once SCEV proves that a phi is congruent,
1950 : // it's often the head of an IV user cycle that is isomorphic
1951 : // with the original phi. It's worth eagerly cleaning up the
1952 : // common case of a single IV increment so that DeleteDeadPHIs
1953 : // can remove cycles that had postinc uses.
1954 : const SCEV *TruncExpr =
1955 57 : SE.getTruncateOrNoop(SE.getSCEV(OrigInc), IsomorphicInc->getType());
1956 55 : if (OrigInc != IsomorphicInc &&
1957 110 : TruncExpr == SE.getSCEV(IsomorphicInc) &&
1958 166 : SE.LI.replacementPreservesLCSSAForm(IsomorphicInc, OrigInc) &&
1959 54 : hoistIVInc(OrigInc, IsomorphicInc)) {
1960 : DEBUG_WITH_TYPE(DebugType,
1961 : dbgs() << "INDVARS: Eliminated congruent iv.inc: "
1962 : << *IsomorphicInc << '\n');
1963 : Value *NewInc = OrigInc;
1964 43 : if (OrigInc->getType() != IsomorphicInc->getType()) {
1965 : Instruction *IP = nullptr;
1966 : if (PHINode *PN = dyn_cast<PHINode>(OrigInc))
1967 1 : IP = &*PN->getParent()->getFirstInsertionPt();
1968 : else
1969 : IP = OrigInc->getNextNode();
1970 :
1971 20 : IRBuilder<> Builder(IP);
1972 20 : Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc());
1973 40 : NewInc = Builder.CreateTruncOrBitCast(
1974 : OrigInc, IsomorphicInc->getType(), IVName);
1975 : }
1976 43 : IsomorphicInc->replaceAllUsesWith(NewInc);
1977 43 : DeadInsts.emplace_back(IsomorphicInc);
1978 : }
1979 : }
1980 : }
1981 : DEBUG_WITH_TYPE(DebugType, dbgs() << "INDVARS: Eliminated congruent iv: "
1982 : << *Phi << '\n');
1983 59 : ++NumElim;
1984 59 : Value *NewIV = OrigPhiRef;
1985 59 : if (OrigPhiRef->getType() != Phi->getType()) {
1986 24 : IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt());
1987 24 : Builder.SetCurrentDebugLocation(Phi->getDebugLoc());
1988 48 : NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName);
1989 : }
1990 59 : Phi->replaceAllUsesWith(NewIV);
1991 59 : DeadInsts.emplace_back(Phi);
1992 : }
1993 13637 : return NumElim;
1994 : }
1995 :
1996 0 : Value *SCEVExpander::getExactExistingExpansion(const SCEV *S,
1997 : const Instruction *At, Loop *L) {
1998 : Optional<ScalarEvolution::ValueOffsetPair> VO =
1999 0 : getRelatedExistingExpansion(S, At, L);
2000 0 : if (VO && VO.getValue().second == nullptr)
2001 0 : return VO.getValue().first;
2002 : return nullptr;
2003 : }
2004 :
2005 : Optional<ScalarEvolution::ValueOffsetPair>
2006 2602 : SCEVExpander::getRelatedExistingExpansion(const SCEV *S, const Instruction *At,
2007 : Loop *L) {
2008 : using namespace llvm::PatternMatch;
2009 :
2010 : SmallVector<BasicBlock *, 4> ExitingBlocks;
2011 2602 : L->getExitingBlocks(ExitingBlocks);
2012 :
2013 : // Look for suitable value in simple conditions at the loop exits.
2014 5685 : for (BasicBlock *BB : ExitingBlocks) {
2015 : ICmpInst::Predicate Pred;
2016 : Instruction *LHS, *RHS;
2017 : BasicBlock *TrueBB, *FalseBB;
2018 :
2019 3103 : if (!match(BB->getTerminator(),
2020 3103 : m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)),
2021 : TrueBB, FalseBB)))
2022 2386 : continue;
2023 :
2024 717 : if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At))
2025 1 : return ScalarEvolution::ValueOffsetPair(LHS, nullptr);
2026 :
2027 716 : if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At))
2028 19 : return ScalarEvolution::ValueOffsetPair(RHS, nullptr);
2029 : }
2030 :
2031 : // Use expand's logic which is used for reusing a previous Value in
2032 : // ExprValueMap.
2033 2582 : ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, At);
2034 2582 : if (VO.first)
2035 : return VO;
2036 :
2037 : // There is potential to make this significantly smarter, but this simple
2038 : // heuristic already gets some interesting cases.
2039 :
2040 : // Can not find suitable value.
2041 : return None;
2042 : }
2043 :
2044 6235 : bool SCEVExpander::isHighCostExpansionHelper(
2045 : const SCEV *S, Loop *L, const Instruction *At,
2046 : SmallPtrSetImpl<const SCEV *> &Processed) {
2047 :
2048 : // If we can find an existing value for this scev available at the point "At"
2049 : // then consider the expression cheap.
2050 8717 : if (At && getRelatedExistingExpansion(S, At, L))
2051 : return false;
2052 :
2053 : // Zero/One operand expressions
2054 11796 : switch (S->getSCEVType()) {
2055 : case scUnknown:
2056 : case scConstant:
2057 : return false;
2058 17 : case scTruncate:
2059 17 : return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(),
2060 17 : L, At, Processed);
2061 10 : case scZeroExtend:
2062 10 : return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(),
2063 10 : L, At, Processed);
2064 11 : case scSignExtend:
2065 11 : return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(),
2066 11 : L, At, Processed);
2067 : }
2068 :
2069 2351 : if (!Processed.insert(S).second)
2070 : return false;
2071 :
2072 : if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) {
2073 : // If the divisor is a power of two and the SCEV type fits in a native
2074 : // integer, consider the division cheap irrespective of whether it occurs in
2075 : // the user code since it can be lowered into a right shift.
2076 540 : if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS()))
2077 537 : if (SC->getAPInt().isPowerOf2()) {
2078 : const DataLayout &DL =
2079 420 : L->getHeader()->getParent()->getParent()->getDataLayout();
2080 : unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth();
2081 420 : return DL.isIllegalInteger(Width);
2082 : }
2083 :
2084 : // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or
2085 : // HowManyLessThans produced to compute a precise expression, rather than a
2086 : // UDiv from the user's code. If we can't find a UDiv in the code with some
2087 : // simple searching, assume the former consider UDivExpr expensive to
2088 : // compute.
2089 120 : BasicBlock *ExitingBB = L->getExitingBlock();
2090 120 : if (!ExitingBB)
2091 : return true;
2092 :
2093 : // At the beginning of this function we already tried to find existing value
2094 : // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern
2095 : // involving division. This is just a simple search heuristic.
2096 120 : if (!At)
2097 : At = &ExitingBB->back();
2098 120 : if (!getRelatedExistingExpansion(
2099 240 : SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L))
2100 : return true;
2101 : }
2102 :
2103 : // HowManyLessThans uses a Max expression whenever the loop is not guarded by
2104 : // the exit condition.
2105 1812 : if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
2106 : return true;
2107 :
2108 : // Recurse past nary expressions, which commonly occur in the
2109 : // BackedgeTakenCount. They may already exist in program code, and if not,
2110 : // they are not too expensive rematerialize.
2111 : if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) {
2112 4924 : for (auto *Op : NAry->operands())
2113 3540 : if (isHighCostExpansionHelper(Op, L, At, Processed))
2114 : return true;
2115 : }
2116 :
2117 : // If we haven't recognized an expensive SCEV pattern, assume it's an
2118 : // expression produced by program code.
2119 : return false;
2120 : }
2121 :
2122 1040 : Value *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred,
2123 : Instruction *IP) {
2124 : assert(IP);
2125 1040 : switch (Pred->getKind()) {
2126 : case SCEVPredicate::P_Union:
2127 948 : return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP);
2128 : case SCEVPredicate::P_Equal:
2129 17 : return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP);
2130 : case SCEVPredicate::P_Wrap: {
2131 : auto *AddRecPred = cast<SCEVWrapPredicate>(Pred);
2132 75 : return expandWrapPredicate(AddRecPred, IP);
2133 : }
2134 : }
2135 0 : llvm_unreachable("Unknown SCEV predicate type");
2136 : }
2137 :
2138 17 : Value *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred,
2139 : Instruction *IP) {
2140 17 : Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP);
2141 17 : Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP);
2142 :
2143 17 : Builder.SetInsertPoint(IP);
2144 17 : auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check");
2145 17 : return I;
2146 : }
2147 :
2148 75 : Value *SCEVExpander::generateOverflowCheck(const SCEVAddRecExpr *AR,
2149 : Instruction *Loc, bool Signed) {
2150 : assert(AR->isAffine() && "Cannot generate RT check for "
2151 : "non-affine expression");
2152 :
2153 75 : SCEVUnionPredicate Pred;
2154 : const SCEV *ExitCount =
2155 75 : SE.getPredicatedBackedgeTakenCount(AR->getLoop(), Pred);
2156 :
2157 : assert(ExitCount != SE.getCouldNotCompute() && "Invalid loop count");
2158 :
2159 75 : const SCEV *Step = AR->getStepRecurrence(SE);
2160 75 : const SCEV *Start = AR->getStart();
2161 :
2162 : Type *ARTy = AR->getType();
2163 75 : unsigned SrcBits = SE.getTypeSizeInBits(ExitCount->getType());
2164 75 : unsigned DstBits = SE.getTypeSizeInBits(ARTy);
2165 :
2166 : // The expression {Start,+,Step} has nusw/nssw if
2167 : // Step < 0, Start - |Step| * Backedge <= Start
2168 : // Step >= 0, Start + |Step| * Backedge > Start
2169 : // and |Step| * Backedge doesn't unsigned overflow.
2170 :
2171 75 : IntegerType *CountTy = IntegerType::get(Loc->getContext(), SrcBits);
2172 75 : Builder.SetInsertPoint(Loc);
2173 75 : Value *TripCountVal = expandCodeFor(ExitCount, CountTy, Loc);
2174 :
2175 : IntegerType *Ty =
2176 75 : IntegerType::get(Loc->getContext(), SE.getTypeSizeInBits(ARTy));
2177 75 : Type *ARExpandTy = DL.isNonIntegralPointerType(ARTy) ? ARTy : Ty;
2178 :
2179 75 : Value *StepValue = expandCodeFor(Step, Ty, Loc);
2180 75 : Value *NegStepValue = expandCodeFor(SE.getNegativeSCEV(Step), Ty, Loc);
2181 75 : Value *StartValue = expandCodeFor(Start, ARExpandTy, Loc);
2182 :
2183 : ConstantInt *Zero =
2184 75 : ConstantInt::get(Loc->getContext(), APInt::getNullValue(DstBits));
2185 :
2186 75 : Builder.SetInsertPoint(Loc);
2187 : // Compute |Step|
2188 150 : Value *StepCompare = Builder.CreateICmp(ICmpInst::ICMP_SLT, StepValue, Zero);
2189 75 : Value *AbsStep = Builder.CreateSelect(StepCompare, NegStepValue, StepValue);
2190 :
2191 : // Get the backedge taken count and truncate or extended to the AR type.
2192 75 : Value *TruncTripCount = Builder.CreateZExtOrTrunc(TripCountVal, Ty);
2193 75 : auto *MulF = Intrinsic::getDeclaration(Loc->getModule(),
2194 150 : Intrinsic::umul_with_overflow, Ty);
2195 :
2196 : // Compute |Step| * Backedge
2197 75 : CallInst *Mul = Builder.CreateCall(MulF, {AbsStep, TruncTripCount}, "mul");
2198 150 : Value *MulV = Builder.CreateExtractValue(Mul, 0, "mul.result");
2199 150 : Value *OfMul = Builder.CreateExtractValue(Mul, 1, "mul.overflow");
2200 :
2201 : // Compute:
2202 : // Start + |Step| * Backedge < Start
2203 : // Start - |Step| * Backedge > Start
2204 : Value *Add = nullptr, *Sub = nullptr;
2205 : if (PointerType *ARPtrTy = dyn_cast<PointerType>(ARExpandTy)) {
2206 2 : const SCEV *MulS = SE.getSCEV(MulV);
2207 2 : const SCEV *NegMulS = SE.getNegativeSCEV(MulS);
2208 4 : Add = Builder.CreateBitCast(expandAddToGEP(MulS, ARPtrTy, Ty, StartValue),
2209 : ARPtrTy);
2210 4 : Sub = Builder.CreateBitCast(
2211 : expandAddToGEP(NegMulS, ARPtrTy, Ty, StartValue), ARPtrTy);
2212 : } else {
2213 73 : Add = Builder.CreateAdd(StartValue, MulV);
2214 73 : Sub = Builder.CreateSub(StartValue, MulV);
2215 : }
2216 :
2217 116 : Value *EndCompareGT = Builder.CreateICmp(
2218 : Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, Sub, StartValue);
2219 :
2220 116 : Value *EndCompareLT = Builder.CreateICmp(
2221 : Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, Add, StartValue);
2222 :
2223 : // Select the answer based on the sign of Step.
2224 : Value *EndCheck =
2225 75 : Builder.CreateSelect(StepCompare, EndCompareGT, EndCompareLT);
2226 :
2227 : // If the backedge taken count type is larger than the AR type,
2228 : // check that we don't drop any bits by truncating it. If we are
2229 : // dropping bits, then we have overflow (unless the step is zero).
2230 75 : if (SE.getTypeSizeInBits(CountTy) > SE.getTypeSizeInBits(Ty)) {
2231 98 : auto MaxVal = APInt::getMaxValue(DstBits).zext(SrcBits);
2232 : auto *BackedgeCheck =
2233 49 : Builder.CreateICmp(ICmpInst::ICMP_UGT, TripCountVal,
2234 49 : ConstantInt::get(Loc->getContext(), MaxVal));
2235 49 : BackedgeCheck = Builder.CreateAnd(
2236 : BackedgeCheck, Builder.CreateICmp(ICmpInst::ICMP_NE, StepValue, Zero));
2237 :
2238 49 : EndCheck = Builder.CreateOr(EndCheck, BackedgeCheck);
2239 : }
2240 :
2241 75 : EndCheck = Builder.CreateOr(EndCheck, OfMul);
2242 75 : return EndCheck;
2243 : }
2244 :
2245 75 : Value *SCEVExpander::expandWrapPredicate(const SCEVWrapPredicate *Pred,
2246 : Instruction *IP) {
2247 75 : const auto *A = cast<SCEVAddRecExpr>(Pred->getExpr());
2248 : Value *NSSWCheck = nullptr, *NUSWCheck = nullptr;
2249 :
2250 : // Add a check for NUSW
2251 75 : if (Pred->getFlags() & SCEVWrapPredicate::IncrementNUSW)
2252 41 : NUSWCheck = generateOverflowCheck(A, IP, false);
2253 :
2254 : // Add a check for NSSW
2255 75 : if (Pred->getFlags() & SCEVWrapPredicate::IncrementNSSW)
2256 34 : NSSWCheck = generateOverflowCheck(A, IP, true);
2257 :
2258 75 : if (NUSWCheck && NSSWCheck)
2259 0 : return Builder.CreateOr(NUSWCheck, NSSWCheck);
2260 :
2261 75 : if (NUSWCheck)
2262 : return NUSWCheck;
2263 :
2264 34 : if (NSSWCheck)
2265 : return NSSWCheck;
2266 :
2267 0 : return ConstantInt::getFalse(IP->getContext());
2268 : }
2269 :
2270 948 : Value *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union,
2271 : Instruction *IP) {
2272 948 : auto *BoolType = IntegerType::get(IP->getContext(), 1);
2273 948 : Value *Check = ConstantInt::getNullValue(BoolType);
2274 :
2275 : // Loop over all checks in this set.
2276 1040 : for (auto Pred : Union->getPredicates()) {
2277 92 : auto *NextCheck = expandCodeForPredicate(Pred, IP);
2278 92 : Builder.SetInsertPoint(IP);
2279 184 : Check = Builder.CreateOr(Check, NextCheck);
2280 : }
2281 :
2282 948 : return Check;
2283 : }
2284 :
2285 : namespace {
2286 : // Search for a SCEV subexpression that is not safe to expand. Any expression
2287 : // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely
2288 : // UDiv expressions. We don't know if the UDiv is derived from an IR divide
2289 : // instruction, but the important thing is that we prove the denominator is
2290 : // nonzero before expansion.
2291 : //
2292 : // IVUsers already checks that IV-derived expressions are safe. So this check is
2293 : // only needed when the expression includes some subexpression that is not IV
2294 : // derived.
2295 : //
2296 : // Currently, we only allow division by a nonzero constant here. If this is
2297 : // inadequate, we could easily allow division by SCEVUnknown by using
2298 : // ValueTracking to check isKnownNonZero().
2299 : //
2300 : // We cannot generally expand recurrences unless the step dominates the loop
2301 : // header. The expander handles the special case of affine recurrences by
2302 : // scaling the recurrence outside the loop, but this technique isn't generally
2303 : // applicable. Expanding a nested recurrence outside a loop requires computing
2304 : // binomial coefficients. This could be done, but the recurrence has to be in a
2305 : // perfectly reduced form, which can't be guaranteed.
2306 : struct SCEVFindUnsafe {
2307 : ScalarEvolution &SE;
2308 : bool IsUnsafe;
2309 :
2310 14375 : SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {}
2311 :
2312 0 : bool follow(const SCEV *S) {
2313 : if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
2314 0 : const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS());
2315 0 : if (!SC || SC->getValue()->isZero()) {
2316 0 : IsUnsafe = true;
2317 0 : return false;
2318 : }
2319 : }
2320 : if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2321 0 : const SCEV *Step = AR->getStepRecurrence(SE);
2322 0 : if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) {
2323 0 : IsUnsafe = true;
2324 0 : return false;
2325 : }
2326 : }
2327 : return true;
2328 : }
2329 0 : bool isDone() const { return IsUnsafe; }
2330 : };
2331 : }
2332 :
2333 : namespace llvm {
2334 14375 : bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) {
2335 : SCEVFindUnsafe Search(SE);
2336 14375 : visitAll(S, Search);
2337 14375 : return !Search.IsUnsafe;
2338 : }
2339 :
2340 189 : bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint,
2341 : ScalarEvolution &SE) {
2342 189 : return isSafeToExpand(S, SE) && SE.dominates(S, InsertionPoint->getParent());
2343 : }
2344 : }
|
